@camstack/addon-vision 0.1.0

package/dist/chunk-LPI42WL6.mjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/shared/node-engine.ts","../src/shared/python-engine.ts","../src/shared/engine-resolver.ts"],"sourcesContent":["import type { IInferenceEngine, DetectionRuntime, DetectionDevice } from '@camstack/types'\nimport * as path from 'node:path'\n\nconst BACKEND_TO_PROVIDER: Readonly<Record<string, string>> = {\n  cpu: 'cpu',\n  coreml: 'coreml',\n  cuda: 'cuda',\n  tensorrt: 'tensorrt',\n  dml: 'dml',\n} as const\n\nconst BACKEND_TO_DEVICE: Readonly<Record<string, DetectionDevice>> = {\n  cpu: 'cpu',\n  coreml: 'gpu-mps',\n  cuda: 'gpu-cuda',\n  tensorrt: 'tensorrt',\n} as const\n\nexport class NodeInferenceEngine implements IInferenceEngine {\n  readonly runtime: DetectionRuntime = 'onnx'\n  readonly device: DetectionDevice\n  private session: unknown = null\n\n  constructor(\n    private readonly modelPath: string,\n    private readonly backend: string,\n  ) {\n    this.device = (BACKEND_TO_DEVICE[backend] ?? 'cpu') as DetectionDevice\n  }\n\n  async initialize(): Promise<void> {\n    const ort = await import('onnxruntime-node')\n    const provider = BACKEND_TO_PROVIDER[this.backend] ?? 'cpu'\n\n    // Resolve absolute path\n    const absModelPath = path.isAbsolute(this.modelPath)\n      ? this.modelPath\n      : path.resolve(process.cwd(), this.modelPath)\n\n    const sessionOptions: Record<string, unknown> = {\n      executionProviders: [provider],\n    }\n\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    this.session = await (ort as any).InferenceSession.create(absModelPath, sessionOptions)\n  }\n\n  async run(input: Float32Array, inputShape: readonly number[]): Promise<Float32Array> {\n    if (!this.session) {\n      throw new Error('NodeInferenceEngine: not initialized — call initialize() first')\n    }\n\n    const ort = await import('onnxruntime-node')\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    const sess = this.session as any\n\n    // Get the first input name\n    const inputName: string = sess.inputNames[0]\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    const tensor = new (ort as any).Tensor('float32', input, [...inputShape])\n    const feeds: Record<string, unknown> = { [inputName]: tensor }\n\n    const results = await sess.run(feeds)\n    const outputName: string = sess.outputNames[0]\n    const outputTensor = results[outputName]\n\n    return outputTensor.data as Float32Array\n  }\n\n  async runMultiOutput(\n    input: Float32Array,\n    inputShape: readonly number[],\n  ): Promise<Record<string, Float32Array>> {\n    if (!this.session) {\n      throw new Error('NodeInferenceEngine: not initialized — call initialize() first')\n    }\n\n    const ort = await import('onnxruntime-node')\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    const sess = this.session as any\n\n    const inputName: string = sess.inputNames[0]\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    const tensor = new (ort as any).Tensor('float32', input, [...inputShape])\n    const feeds: Record<string, unknown> = { [inputName]: tensor }\n\n    const results = await sess.run(feeds)\n\n    const out: Record<string, Float32Array> = {}\n    for (const name of sess.outputNames as string[]) {\n      out[name] = results[name].data as Float32Array\n    }\n    return out\n  }\n\n  async dispose(): Promise<void> {\n    // onnxruntime-node sessions don't have explicit close in all versions\n    // but we clear the reference\n    this.session = null\n  }\n}\n","import type { IInferenceEngine, DetectionRuntime, DetectionDevice } from '@camstack/types'\nimport { spawn, type ChildProcess } from 'node:child_process'\n\nexport class PythonInferenceEngine implements IInferenceEngine {\n  readonly runtime: DetectionRuntime\n  readonly device: DetectionDevice\n  private process: ChildProcess | null = null\n  private receiveBuffer: Buffer = Buffer.alloc(0)\n  private pendingResolve: ((value: Record<string, unknown>) => void) | null = null\n  private pendingReject: ((reason: Error) => void) | null = null\n\n  constructor(\n    private readonly pythonPath: string,\n    private readonly scriptPath: string,\n    runtime: DetectionRuntime,\n    private readonly modelPath: string,\n    private readonly extraArgs: readonly string[] = [],\n  ) {\n    this.runtime = runtime\n    // Determine device from runtime\n    const runtimeDeviceMap: Readonly<Record<DetectionRuntime, DetectionDevice>> = {\n      onnx: 'cpu',\n      coreml: 'gpu-mps',\n      pytorch: 'cpu',\n      openvino: 'cpu',\n      tflite: 'cpu',\n    }\n    this.device = runtimeDeviceMap[runtime]\n  }\n\n  async initialize(): Promise<void> {\n    const args = [this.scriptPath, this.modelPath, ...this.extraArgs]\n    this.process = spawn(this.pythonPath, args, {\n      stdio: ['pipe', 'pipe', 'pipe'],\n    })\n\n    if (!this.process.stdout || !this.process.stdin) {\n      throw new Error('PythonInferenceEngine: failed to create process pipes')\n    }\n\n    this.process.stderr?.on('data', (chunk: Buffer) => {\n      // Log stderr from python process for debugging\n      process.stderr.write(`[python-engine] ${chunk.toString()}`)\n    })\n\n    this.process.on('error', (err) => {\n      this.pendingReject?.(err)\n      this.pendingReject = null\n      this.pendingResolve = null\n    })\n\n    this.process.on('exit', (code) => {\n      if (code !== 0) {\n        const err = new Error(`PythonInferenceEngine: process exited with code ${code}`)\n        this.pendingReject?.(err)\n        this.pendingReject = null\n        this.pendingResolve = null\n      }\n    })\n\n    this.process.stdout.on('data', (chunk: Buffer) => {\n      this.receiveBuffer = Buffer.concat([this.receiveBuffer, chunk])\n      this._tryReceive()\n    })\n\n    // Give the process a moment to start up and load the model\n    await new Promise<void>((resolve, reject) => {\n      const timeout = setTimeout(() => resolve(), 2000)\n      this.process?.on('error', (err) => {\n        clearTimeout(timeout)\n        reject(err)\n      })\n      this.process?.on('exit', (code) => {\n        clearTimeout(timeout)\n        if (code !== 0) {\n          reject(new Error(`PythonInferenceEngine: process exited early with code ${code}`))\n        }\n      })\n    })\n  }\n\n  private _tryReceive(): void {\n    // Binary IPC: [4 bytes LE uint32 length][JSON bytes]\n    if (this.receiveBuffer.length < 4) return\n\n    const length = this.receiveBuffer.readUInt32LE(0)\n    if (this.receiveBuffer.length < 4 + length) return\n\n    const jsonBytes = this.receiveBuffer.subarray(4, 4 + length)\n    this.receiveBuffer = this.receiveBuffer.subarray(4 + length)\n\n    const resolve = this.pendingResolve\n    const reject = this.pendingReject\n    this.pendingResolve = null\n    this.pendingReject = null\n\n    if (!resolve) return\n\n    try {\n      const parsed = JSON.parse(jsonBytes.toString('utf8')) as Record<string, unknown>\n      resolve(parsed)\n    } catch (err) {\n      reject?.(err instanceof Error ? err : new Error(String(err)))\n    }\n  }\n\n  /** Send JPEG buffer, receive JSON detection results */\n  async runJpeg(jpeg: Buffer): Promise<Record<string, unknown>> {\n    if (!this.process?.stdin) {\n      throw new Error('PythonInferenceEngine: process not initialized')\n    }\n\n    return new Promise<Record<string, unknown>>((resolve, reject) => {\n      this.pendingResolve = resolve\n      this.pendingReject = reject\n\n      // Binary IPC: [4 bytes LE uint32 length][JPEG bytes]\n      const lengthBuf = Buffer.allocUnsafe(4)\n      lengthBuf.writeUInt32LE(jpeg.length, 0)\n      this.process!.stdin!.write(Buffer.concat([lengthBuf, jpeg]))\n    })\n  }\n\n  /** IInferenceEngine.run — wraps runJpeg for compatibility */\n  async run(_input: Float32Array, _inputShape: readonly number[]): Promise<Float32Array> {\n    throw new Error(\n      'PythonInferenceEngine: use runJpeg() directly — this engine operates on JPEG input',\n    )\n  }\n\n  /** IInferenceEngine.runMultiOutput — not supported by Python engine (operates on JPEG input) */\n  async runMultiOutput(\n    _input: Float32Array,\n    _inputShape: readonly number[],\n  ): Promise<Record<string, Float32Array>> {\n    throw new Error(\n      'PythonInferenceEngine: runMultiOutput() is not supported — this engine operates on JPEG input',\n    )\n  }\n\n  async dispose(): Promise<void> {\n    if (this.process) {\n      this.process.stdin?.end()\n      this.process.kill('SIGTERM')\n      this.process = null\n    }\n  }\n}\n","// TODO: Wire PythonInferenceEngine for PyTorch/OpenVINO/TFLite runtimes\n// Currently falls back to ONNX CPU when non-ONNX runtime is requested.\n// See: packages/addon-vision/python/ for stub implementations.\n//\n// WHY THIS FILE USES RAW FILESYSTEM PATHS (modelsDir: string):\n//\n// Model files must be loaded via absolute filesystem paths because inference\n// engines (ONNX Runtime, CoreML, etc.) require direct file access — they do\n// not accept Buffer or stream inputs. This is fundamentally different from\n// user data (recordings, media) where IAddonFileStorage abstraction makes\n// sense. Models are closer to binaries or compiled artifacts that must live\n// on the local filesystem at runtime.\n//\n// IAddonFileStorage.readFile() returns a Buffer, but onnxruntime-node's\n// InferenceSession.create() only accepts a file path string or a Uint8Array\n// loaded into memory. For large models (hundreds of MB), loading into a\n// Uint8Array is impractical. Therefore, modelsDir stays as a raw string path\n// and is intentionally NOT replaced with IAddonFileStorage here.\n\nimport type {\n  IInferenceEngine,\n  DetectionRuntime,\n  ModelCatalogEntry,\n  ModelFormat,\n} from '@camstack/types'\nimport * as fs from 'node:fs'\nimport * as path from 'node:path'\nimport { NodeInferenceEngine } from './node-engine.js'\nimport { PythonInferenceEngine } from './python-engine.js'\n\nexport interface EngineResolverOptions {\n  readonly runtime: DetectionRuntime | 'auto'\n  readonly backend: string\n  readonly modelEntry: ModelCatalogEntry\n  readonly modelsDir: string\n  readonly pythonPath?: string\n  readonly downloadModel?: (url: string, destDir: string) => Promise<string>\n}\n\nexport interface ResolvedEngine {\n  readonly engine: IInferenceEngine\n  readonly format: ModelFormat\n  readonly modelPath: string\n}\n\n/** Priority order for auto-selection of ONNX backends */\nconst AUTO_BACKEND_PRIORITY = ['coreml', 'cuda', 'tensorrt', 'cpu'] as const\n\n/** Map backend names to the model format they require */\nconst BACKEND_TO_FORMAT: Readonly<Record<string, ModelFormat>> = {\n  cpu: 'onnx',\n  coreml: 'coreml',\n  cuda: 'onnx',\n  tensorrt: 'onnx',\n} as const\n\n/** Map DetectionRuntime to ModelFormat */\nconst RUNTIME_TO_FORMAT: Readonly<Partial<Record<DetectionRuntime, ModelFormat>>> = {\n  onnx: 'onnx',\n  coreml: 'coreml',\n  openvino: 'openvino',\n  tflite: 'tflite',\n  pytorch: 'pt',\n} as const\n\nfunction modelFilePath(modelsDir: string, modelEntry: ModelCatalogEntry, format: ModelFormat): string {\n  const formatEntry = modelEntry.formats[format]\n  if (!formatEntry) {\n    throw new Error(`Model ${modelEntry.id} has no ${format} format`)\n  }\n  // Derive filename from URL\n  const urlParts = formatEntry.url.split('/')\n  const filename = urlParts[urlParts.length - 1] ?? `${modelEntry.id}.${format}`\n  return path.join(modelsDir, filename)\n}\n\nfunction modelExists(filePath: string): boolean {\n  try {\n    return fs.existsSync(filePath)\n  } catch {\n    return false\n  }\n}\n\nexport async function resolveEngine(options: EngineResolverOptions): Promise<ResolvedEngine> {\n  const { runtime, backend, modelEntry, modelsDir, downloadModel } = options\n\n  let selectedFormat: ModelFormat\n  let selectedBackend: string\n\n  if (runtime === 'auto') {\n    // Probe available ONNX backends and pick best\n    const available = await probeOnnxBackends()\n\n    // Pick first priority backend that has a corresponding model format available\n    let chosen: { backend: string; format: ModelFormat } | null = null\n\n    for (const b of AUTO_BACKEND_PRIORITY) {\n      if (!available.includes(b)) continue\n      const fmt = BACKEND_TO_FORMAT[b]\n      if (!fmt) continue\n      if (!modelEntry.formats[fmt]) continue\n      chosen = { backend: b, format: fmt }\n      break\n    }\n\n    if (!chosen) {\n      throw new Error(\n        `resolveEngine: no compatible backend found for model ${modelEntry.id}. Available backends: ${available.join(', ')}`,\n      )\n    }\n\n    selectedFormat = chosen.format\n    selectedBackend = chosen.backend\n  } else {\n    // Explicit runtime requested\n    const fmt = RUNTIME_TO_FORMAT[runtime]\n    if (!fmt) {\n      throw new Error(`resolveEngine: unsupported runtime \"${runtime}\"`)\n    }\n    if (!modelEntry.formats[fmt]) {\n      throw new Error(\n        `resolveEngine: model ${modelEntry.id} has no ${fmt} format for runtime ${runtime}`,\n      )\n    }\n    selectedFormat = fmt\n    // For onnx runtime, use the provided backend; otherwise use the runtime name\n    selectedBackend = runtime === 'onnx' ? (backend || 'cpu') : runtime\n  }\n\n  // Resolve model path\n  let modelPath = modelFilePath(modelsDir, modelEntry, selectedFormat)\n\n  if (!modelExists(modelPath)) {\n    if (downloadModel) {\n      const formatEntry = modelEntry.formats[selectedFormat]!\n      modelPath = await downloadModel(formatEntry.url, modelsDir)\n    } else {\n      throw new Error(\n        `resolveEngine: model file not found at ${modelPath} and no downloadModel function provided`,\n      )\n    }\n  }\n\n  // Only ONNX runtime is handled by NodeInferenceEngine currently\n  if (selectedFormat === 'onnx' || selectedFormat === 'coreml') {\n    const engine = new NodeInferenceEngine(modelPath, selectedBackend)\n    await engine.initialize()\n    return { engine, format: selectedFormat, modelPath }\n  }\n\n  // For non-ONNX/CoreML formats, try PythonInferenceEngine when a python binary is available\n  const { pythonPath } = options\n  const PYTHON_SCRIPT_MAP: Readonly<Record<string, string>> = {\n    coreml: 'coreml_inference.py',\n    pytorch: 'pytorch_inference.py',\n    openvino: 'openvino_inference.py',\n  } as const\n\n  const effectiveRuntime = runtime === 'auto' ? selectedBackend : runtime\n  const scriptName = PYTHON_SCRIPT_MAP[effectiveRuntime]\n\n  if (scriptName && pythonPath) {\n    const scriptPath = path.join(__dirname, '../../python', scriptName)\n    const inputSize = Math.max(modelEntry.inputSize.width, modelEntry.inputSize.height)\n    const engine = new PythonInferenceEngine(pythonPath, scriptPath, effectiveRuntime as DetectionRuntime, modelPath, [\n      `--input-size=${inputSize}`,\n      `--confidence=0.25`,\n    ])\n    await engine.initialize()\n    return { engine, format: selectedFormat, modelPath }\n  }\n\n  // Final fallback: use ONNX CPU if available\n  const fallbackPath = modelFilePath(modelsDir, modelEntry, 'onnx')\n  if (modelEntry.formats['onnx'] && modelExists(fallbackPath)) {\n    const engine = new NodeInferenceEngine(fallbackPath, 'cpu')\n    await engine.initialize()\n    return { engine, format: 'onnx', modelPath: fallbackPath }\n  }\n\n  throw new Error(\n    `resolveEngine: format ${selectedFormat} is not yet supported by NodeInferenceEngine, ` +\n    `no Python runtime is available, and no ONNX fallback exists`,\n  )\n}\n\n/** Probe which ONNX execution providers are available on this system */\nexport async function probeOnnxBackends(): Promise<string[]> {\n  const available: string[] = ['cpu'] // CPU is always available\n\n  try {\n    const ort = await import('onnxruntime-node')\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    const providers: string[] = (ort as any).env?.webgl?.disabled !== undefined\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      ? ((ort as any).InferenceSession?.getAvailableProviders?.() ?? [])\n      : []\n\n    for (const p of providers) {\n      const normalized = p.toLowerCase().replace('executionprovider', '')\n      if (normalized === 'coreml') available.push('coreml')\n      else if (normalized === 'cuda') available.push('cuda')\n      else if (normalized === 'tensorrt') available.push('tensorrt')\n    }\n  } catch {\n    // onnxruntime-node may not be installed; CPU only\n  }\n\n  // Platform-specific hints when getAvailableProviders isn't exposed\n  if (process.platform === 'darwin' && !available.includes('coreml')) {\n    available.push('coreml')\n  }\n\n  return [...new Set(available)]\n}\n"],"mappings":";AACA,YAAY,UAAU;AAEtB,IAAM,sBAAwD;AAAA,EAC5D,KAAK;AAAA,EACL,QAAQ;AAAA,EACR,MAAM;AAAA,EACN,UAAU;AAAA,EACV,KAAK;AACP;AAEA,IAAM,oBAA+D;AAAA,EACnE,KAAK;AAAA,EACL,QAAQ;AAAA,EACR,MAAM;AAAA,EACN,UAAU;AACZ;AAEO,IAAM,sBAAN,MAAsD;AAAA,EAK3D,YACmB,WACA,SACjB;AAFiB;AACA;AAEjB,SAAK,SAAU,kBAAkB,OAAO,KAAK;AAAA,EAC/C;AAAA,EATS,UAA4B;AAAA,EAC5B;AAAA,EACD,UAAmB;AAAA,EAS3B,MAAM,aAA4B;AAChC,UAAM,MAAM,MAAM,OAAO,kBAAkB;AAC3C,UAAM,WAAW,oBAAoB,KAAK,OAAO,KAAK;AAGtD,UAAM,eAAoB,gBAAW,KAAK,SAAS,IAC/C,KAAK,YACA,aAAQ,QAAQ,IAAI,GAAG,KAAK,SAAS;AAE9C,UAAM,iBAA0C;AAAA,MAC9C,oBAAoB,CAAC,QAAQ;AAAA,IAC/B;AAGA,SAAK,UAAU,MAAO,IAAY,iBAAiB,OAAO,cAAc,cAAc;AAAA,EACxF;AAAA,EAEA,MAAM,IAAI,OAAqB,YAAsD;AACnF,QAAI,CAAC,KAAK,SAAS;AACjB,YAAM,IAAI,MAAM,qEAAgE;AAAA,IAClF;AAEA,UAAM,MAAM,MAAM,OAAO,kBAAkB;AAE3C,UAAM,OAAO,KAAK;AAGlB,UAAM,YAAoB,KAAK,WAAW,CAAC;AAE3C,UAAM,SAAS,IAAK,IAAY,OAAO,WAAW,OAAO,CAAC,GAAG,UAAU,CAAC;AACxE,UAAM,QAAiC,EAAE,CAAC,SAAS,GAAG,OAAO;AAE7D,UAAM,UAAU,MAAM,KAAK,IAAI,KAAK;AACpC,UAAM,aAAqB,KAAK,YAAY,CAAC;AAC7C,UAAM,eAAe,QAAQ,UAAU;AAEvC,WAAO,aAAa;AAAA,EACtB;AAAA,EAEA,MAAM,eACJ,OACA,YACuC;AACvC,QAAI,CAAC,KAAK,SAAS;AACjB,YAAM,IAAI,MAAM,qEAAgE;AAAA,IAClF;AAEA,UAAM,MAAM,MAAM,OAAO,kBAAkB;AAE3C,UAAM,OAAO,KAAK;AAElB,UAAM,YAAoB,KAAK,WAAW,CAAC;AAE3C,UAAM,SAAS,IAAK,IAAY,OAAO,WAAW,OAAO,CAAC,GAAG,UAAU,CAAC;AACxE,UAAM,QAAiC,EAAE,CAAC,SAAS,GAAG,OAAO;AAE7D,UAAM,UAAU,MAAM,KAAK,IAAI,KAAK;AAEpC,UAAM,MAAoC,CAAC;AAC3C,eAAW,QAAQ,KAAK,aAAyB;AAC/C,UAAI,IAAI,IAAI,QAAQ,IAAI,EAAE;AAAA,IAC5B;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,UAAyB;AAG7B,SAAK,UAAU;AAAA,EACjB;AACF;;;ACnGA,SAAS,aAAgC;AAElC,IAAM,wBAAN,MAAwD;AAAA,EAQ7D,YACmB,YACA,YACjB,SACiB,WACA,YAA+B,CAAC,GACjD;AALiB;AACA;AAEA;AACA;AAEjB,SAAK,UAAU;AAEf,UAAM,mBAAwE;AAAA,MAC5E,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,SAAS;AAAA,MACT,UAAU;AAAA,MACV,QAAQ;AAAA,IACV;AACA,SAAK,SAAS,iBAAiB,OAAO;AAAA,EACxC;AAAA,EAxBS;AAAA,EACA;AAAA,EACD,UAA+B;AAAA,EAC/B,gBAAwB,OAAO,MAAM,CAAC;AAAA,EACtC,iBAAoE;AAAA,EACpE,gBAAkD;AAAA,EAqB1D,MAAM,aAA4B;AAChC,UAAM,OAAO,CAAC,KAAK,YAAY,KAAK,WAAW,GAAG,KAAK,SAAS;AAChE,SAAK,UAAU,MAAM,KAAK,YAAY,MAAM;AAAA,MAC1C,OAAO,CAAC,QAAQ,QAAQ,MAAM;AAAA,IAChC,CAAC;AAED,QAAI,CAAC,KAAK,QAAQ,UAAU,CAAC,KAAK,QAAQ,OAAO;AAC/C,YAAM,IAAI,MAAM,uDAAuD;AAAA,IACzE;AAEA,SAAK,QAAQ,QAAQ,GAAG,QAAQ,CAAC,UAAkB;AAEjD,cAAQ,OAAO,MAAM,mBAAmB,MAAM,SAAS,CAAC,EAAE;AAAA,IAC5D,CAAC;AAED,SAAK,QAAQ,GAAG,SAAS,CAAC,QAAQ;AAChC,WAAK,gBAAgB,GAAG;AACxB,WAAK,gBAAgB;AACrB,WAAK,iBAAiB;AAAA,IACxB,CAAC;AAED,SAAK,QAAQ,GAAG,QAAQ,CAAC,SAAS;AAChC,UAAI,SAAS,GAAG;AACd,cAAM,MAAM,IAAI,MAAM,mDAAmD,IAAI,EAAE;AAC/E,aAAK,gBAAgB,GAAG;AACxB,aAAK,gBAAgB;AACrB,aAAK,iBAAiB;AAAA,MACxB;AAAA,IACF,CAAC;AAED,SAAK,QAAQ,OAAO,GAAG,QAAQ,CAAC,UAAkB;AAChD,WAAK,gBAAgB,OAAO,OAAO,CAAC,KAAK,eAAe,KAAK,CAAC;AAC9D,WAAK,YAAY;AAAA,IACnB,CAAC;AAGD,UAAM,IAAI,QAAc,CAACA,UAAS,WAAW;AAC3C,YAAM,UAAU,WAAW,MAAMA,SAAQ,GAAG,GAAI;AAChD,WAAK,SAAS,GAAG,SAAS,CAAC,QAAQ;AACjC,qBAAa,OAAO;AACpB,eAAO,GAAG;AAAA,MACZ,CAAC;AACD,WAAK,SAAS,GAAG,QAAQ,CAAC,SAAS;AACjC,qBAAa,OAAO;AACpB,YAAI,SAAS,GAAG;AACd,iBAAO,IAAI,MAAM,yDAAyD,IAAI,EAAE,CAAC;AAAA,QACnF;AAAA,MACF,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AAAA,EAEQ,cAAoB;AAE1B,QAAI,KAAK,cAAc,SAAS,EAAG;AAEnC,UAAM,SAAS,KAAK,cAAc,aAAa,CAAC;AAChD,QAAI,KAAK,cAAc,SAAS,IAAI,OAAQ;AAE5C,UAAM,YAAY,KAAK,cAAc,SAAS,GAAG,IAAI,MAAM;AAC3D,SAAK,gBAAgB,KAAK,cAAc,SAAS,IAAI,MAAM;AAE3D,UAAMA,WAAU,KAAK;AACrB,UAAM,SAAS,KAAK;AACpB,SAAK,iBAAiB;AACtB,SAAK,gBAAgB;AAErB,QAAI,CAACA,SAAS;AAEd,QAAI;AACF,YAAM,SAAS,KAAK,MAAM,UAAU,SAAS,MAAM,CAAC;AACpD,MAAAA,SAAQ,MAAM;AAAA,IAChB,SAAS,KAAK;AACZ,eAAS,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,GAAG,CAAC,CAAC;AAAA,IAC9D;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,QAAQ,MAAgD;AAC5D,QAAI,CAAC,KAAK,SAAS,OAAO;AACxB,YAAM,IAAI,MAAM,gDAAgD;AAAA,IAClE;AAEA,WAAO,IAAI,QAAiC,CAACA,UAAS,WAAW;AAC/D,WAAK,iBAAiBA;AACtB,WAAK,gBAAgB;AAGrB,YAAM,YAAY,OAAO,YAAY,CAAC;AACtC,gBAAU,cAAc,KAAK,QAAQ,CAAC;AACtC,WAAK,QAAS,MAAO,MAAM,OAAO,OAAO,CAAC,WAAW,IAAI,CAAC,CAAC;AAAA,IAC7D,CAAC;AAAA,EACH;AAAA;AAAA,EAGA,MAAM,IAAI,QAAsB,aAAuD;AACrF,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,eACJ,QACA,aACuC;AACvC,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAM,UAAyB;AAC7B,QAAI,KAAK,SAAS;AAChB,WAAK,QAAQ,OAAO,IAAI;AACxB,WAAK,QAAQ,KAAK,SAAS;AAC3B,WAAK,UAAU;AAAA,IACjB;AAAA,EACF;AACF;;;AC1HA,YAAY,QAAQ;AACpB,YAAYC,WAAU;AAoBtB,IAAM,wBAAwB,CAAC,UAAU,QAAQ,YAAY,KAAK;AAGlE,IAAM,oBAA2D;AAAA,EAC/D,KAAK;AAAA,EACL,QAAQ;AAAA,EACR,MAAM;AAAA,EACN,UAAU;AACZ;AAGA,IAAM,oBAA8E;AAAA,EAClF,MAAM;AAAA,EACN,QAAQ;AAAA,EACR,UAAU;AAAA,EACV,QAAQ;AAAA,EACR,SAAS;AACX;AAEA,SAAS,cAAc,WAAmB,YAA+B,QAA6B;AACpG,QAAM,cAAc,WAAW,QAAQ,MAAM;AAC7C,MAAI,CAAC,aAAa;AAChB,UAAM,IAAI,MAAM,SAAS,WAAW,EAAE,WAAW,MAAM,SAAS;AAAA,EAClE;AAEA,QAAM,WAAW,YAAY,IAAI,MAAM,GAAG;AAC1C,QAAM,WAAW,SAAS,SAAS,SAAS,CAAC,KAAK,GAAG,WAAW,EAAE,IAAI,MAAM;AAC5E,SAAY,WAAK,WAAW,QAAQ;AACtC;AAEA,SAAS,YAAY,UAA2B;AAC9C,MAAI;AACF,WAAU,cAAW,QAAQ;AAAA,EAC/B,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,eAAsB,cAAc,SAAyD;AAC3F,QAAM,EAAE,SAAS,SAAS,YAAY,WAAW,cAAc,IAAI;AAEnE,MAAI;AACJ,MAAI;AAEJ,MAAI,YAAY,QAAQ;AAEtB,UAAM,YAAY,MAAM,kBAAkB;AAG1C,QAAI,SAA0D;AAE9D,eAAW,KAAK,uBAAuB;AACrC,UAAI,CAAC,UAAU,SAAS,CAAC,EAAG;AAC5B,YAAM,MAAM,kBAAkB,CAAC;AAC/B,UAAI,CAAC,IAAK;AACV,UAAI,CAAC,WAAW,QAAQ,GAAG,EAAG;AAC9B,eAAS,EAAE,SAAS,GAAG,QAAQ,IAAI;AACnC;AAAA,IACF;AAEA,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR,wDAAwD,WAAW,EAAE,yBAAyB,UAAU,KAAK,IAAI,CAAC;AAAA,MACpH;AAAA,IACF;AAEA,qBAAiB,OAAO;AACxB,sBAAkB,OAAO;AAAA,EAC3B,OAAO;AAEL,UAAM,MAAM,kBAAkB,OAAO;AACrC,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,uCAAuC,OAAO,GAAG;AAAA,IACnE;AACA,QAAI,CAAC,WAAW,QAAQ,GAAG,GAAG;AAC5B,YAAM,IAAI;AAAA,QACR,wBAAwB,WAAW,EAAE,WAAW,GAAG,uBAAuB,OAAO;AAAA,MACnF;AAAA,IACF;AACA,qBAAiB;AAEjB,sBAAkB,YAAY,SAAU,WAAW,QAAS;AAAA,EAC9D;AAGA,MAAI,YAAY,cAAc,WAAW,YAAY,cAAc;AAEnE,MAAI,CAAC,YAAY,SAAS,GAAG;AAC3B,QAAI,eAAe;AACjB,YAAM,cAAc,WAAW,QAAQ,cAAc;AACrD,kBAAY,MAAM,cAAc,YAAY,KAAK,SAAS;AAAA,IAC5D,OAAO;AACL,YAAM,IAAI;AAAA,QACR,0CAA0C,SAAS;AAAA,MACrD;AAAA,IACF;AAAA,EACF;AAGA,MAAI,mBAAmB,UAAU,mBAAmB,UAAU;AAC5D,UAAM,SAAS,IAAI,oBAAoB,WAAW,eAAe;AACjE,UAAM,OAAO,WAAW;AACxB,WAAO,EAAE,QAAQ,QAAQ,gBAAgB,UAAU;AAAA,EACrD;AAGA,QAAM,EAAE,WAAW,IAAI;AACvB,QAAM,oBAAsD;AAAA,IAC1D,QAAQ;AAAA,IACR,SAAS;AAAA,IACT,UAAU;AAAA,EACZ;AAEA,QAAM,mBAAmB,YAAY,SAAS,kBAAkB;AAChE,QAAM,aAAa,kBAAkB,gBAAgB;AAErD,MAAI,cAAc,YAAY;AAC5B,UAAM,aAAkB,WAAK,WAAW,gBAAgB,UAAU;AAClE,UAAM,YAAY,KAAK,IAAI,WAAW,UAAU,OAAO,WAAW,UAAU,MAAM;AAClF,UAAM,SAAS,IAAI,sBAAsB,YAAY,YAAY,kBAAsC,WAAW;AAAA,MAChH,gBAAgB,SAAS;AAAA,MACzB;AAAA,IACF,CAAC;AACD,UAAM,OAAO,WAAW;AACxB,WAAO,EAAE,QAAQ,QAAQ,gBAAgB,UAAU;AAAA,EACrD;AAGA,QAAM,eAAe,cAAc,WAAW,YAAY,MAAM;AAChE,MAAI,WAAW,QAAQ,MAAM,KAAK,YAAY,YAAY,GAAG;AAC3D,UAAM,SAAS,IAAI,oBAAoB,cAAc,KAAK;AAC1D,UAAM,OAAO,WAAW;AACxB,WAAO,EAAE,QAAQ,QAAQ,QAAQ,WAAW,aAAa;AAAA,EAC3D;AAEA,QAAM,IAAI;AAAA,IACR,yBAAyB,cAAc;AAAA,EAEzC;AACF;AAGA,eAAsB,oBAAuC;AAC3D,QAAM,YAAsB,CAAC,KAAK;AAElC,MAAI;AACF,UAAM,MAAM,MAAM,OAAO,kBAAkB;AAE3C,UAAM,YAAuB,IAAY,KAAK,OAAO,aAAa,SAE5D,IAAY,kBAAkB,wBAAwB,KAAK,CAAC,IAC9D,CAAC;AAEL,eAAW,KAAK,WAAW;AACzB,YAAM,aAAa,EAAE,YAAY,EAAE,QAAQ,qBAAqB,EAAE;AAClE,UAAI,eAAe,SAAU,WAAU,KAAK,QAAQ;AAAA,eAC3C,eAAe,OAAQ,WAAU,KAAK,MAAM;AAAA,eAC5C,eAAe,WAAY,WAAU,KAAK,UAAU;AAAA,IAC/D;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,MAAI,QAAQ,aAAa,YAAY,CAAC,UAAU,SAAS,QAAQ,GAAG;AAClE,cAAU,KAAK,QAAQ;AAAA,EACzB;AAEA,SAAO,CAAC,GAAG,IAAI,IAAI,SAAS,CAAC;AAC/B;","names":["resolve","path"]}

package/dist/chunk-MEVASN3P.mjs

@@ -0,0 +1,305 @@
+import {
+  nms
+} from "./chunk-KUO2BVFY.mjs";
+import {
+  cropRegion,
+  letterbox
+} from "./chunk-6OR5TE7A.mjs";
+import {
+  resolveEngine
+} from "./chunk-LPI42WL6.mjs";
+
+// src/catalogs/face-detection-models.ts
+import { hfModelUrl } from "@camstack/types";
+var HF_REPO = "camstack/camstack-models";
+var FACE_LABELS = [
+  { id: "face", name: "Face" }
+];
+var FACE_DETECTION_MODELS = [
+  {
+    id: "scrfd-500m",
+    name: "SCRFD 500M",
+    description: "SCRFD 500M \u2014 ultra-lightweight face detector",
+    inputSize: { width: 640, height: 640 },
+    labels: FACE_LABELS,
+    formats: {
+      onnx: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/onnx/camstack-scrfd-500m.onnx"),
+        sizeMB: 2.2
+      },
+      coreml: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/coreml/camstack-scrfd-500m.mlpackage"),
+        sizeMB: 1.2
+      },
+      openvino: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/openvino/camstack-scrfd-500m.xml"),
+        sizeMB: 1.3
+      }
+    }
+  },
+  {
+    id: "scrfd-2.5g",
+    name: "SCRFD 2.5G",
+    description: "SCRFD 2.5G \u2014 balanced face detection model",
+    inputSize: { width: 640, height: 640 },
+    labels: FACE_LABELS,
+    formats: {
+      onnx: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/onnx/camstack-scrfd-2.5g.onnx"),
+        sizeMB: 3.1
+      },
+      coreml: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/coreml/camstack-scrfd-2.5g.mlpackage"),
+        sizeMB: 1.7
+      },
+      openvino: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/openvino/camstack-scrfd-2.5g.xml"),
+        sizeMB: 1.8
+      }
+    }
+  },
+  {
+    id: "scrfd-10g",
+    name: "SCRFD 10G",
+    description: "SCRFD 10G \u2014 high-accuracy face detector",
+    inputSize: { width: 640, height: 640 },
+    labels: FACE_LABELS,
+    formats: {
+      onnx: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/onnx/camstack-scrfd-10g.onnx"),
+        sizeMB: 16
+      },
+      coreml: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/coreml/camstack-scrfd-10g.mlpackage"),
+        sizeMB: 8.2
+      },
+      openvino: {
+        url: hfModelUrl(HF_REPO, "faceDetection/scrfd/openvino/camstack-scrfd-10g.xml"),
+        sizeMB: 8.3
+      }
+    }
+  }
+];
+
+// src/shared/postprocess/scrfd.ts
+var STRIDES = [8, 16, 32];
+var NUM_ANCHORS_PER_STRIDE = 2;
+function generateAnchors(stride, inputSize) {
+  const featureSize = Math.ceil(inputSize / stride);
+  const anchors = [];
+  for (let y = 0; y < featureSize; y++) {
+    for (let x = 0; x < featureSize; x++) {
+      for (let k = 0; k < NUM_ANCHORS_PER_STRIDE; k++) {
+        anchors.push({
+          cx: (x + 0.5) * stride,
+          cy: (y + 0.5) * stride
+        });
+      }
+    }
+  }
+  return anchors;
+}
+function scrfdPostprocess(outputs, confidence, inputSize, originalWidth, originalHeight) {
+  const scaleX = originalWidth / inputSize;
+  const scaleY = originalHeight / inputSize;
+  const candidates = [];
+  for (const stride of STRIDES) {
+    const scoreKey = Object.keys(outputs).find((k) => k.includes(`score_${stride}`) || k.includes(`_${stride}_score`));
+    const bboxKey = Object.keys(outputs).find((k) => k.includes(`bbox_${stride}`) || k.includes(`_${stride}_bbox`));
+    const kpsKey = Object.keys(outputs).find((k) => k.includes(`kps_${stride}`) || k.includes(`_${stride}_kps`));
+    if (!scoreKey || !bboxKey) continue;
+    const scores = outputs[scoreKey];
+    const bboxes = outputs[bboxKey];
+    const kps = kpsKey ? outputs[kpsKey] : void 0;
+    const anchors = generateAnchors(stride, inputSize);
+    const n = anchors.length;
+    for (let i = 0; i < n; i++) {
+      const score = scores[i];
+      if (score < confidence) continue;
+      const anchor = anchors[i];
+      const x1 = anchor.cx - bboxes[i * 4] * stride;
+      const y1 = anchor.cy - bboxes[i * 4 + 1] * stride;
+      const x2 = anchor.cx + bboxes[i * 4 + 2] * stride;
+      const y2 = anchor.cy + bboxes[i * 4 + 3] * stride;
+      const bbox = {
+        x: x1 * scaleX,
+        y: y1 * scaleY,
+        w: (x2 - x1) * scaleX,
+        h: (y2 - y1) * scaleY
+      };
+      let landmarks;
+      if (kps) {
+        const pts = [];
+        for (let p = 0; p < 5; p++) {
+          pts.push({
+            x: (anchor.cx + kps[i * 10 + p * 2] * stride) * scaleX,
+            y: (anchor.cy + kps[i * 10 + p * 2 + 1] * stride) * scaleY
+          });
+        }
+        landmarks = pts;
+      }
+      candidates.push({ bbox, score, landmarks });
+    }
+  }
+  if (candidates.length === 0) return [];
+  const keptIndices = nms(candidates, 0.45);
+  return keptIndices.map((idx) => {
+    const { bbox, score, landmarks } = candidates[idx];
+    return {
+      class: "face",
+      originalClass: "face",
+      score,
+      bbox,
+      ...landmarks ? { landmarks } : {}
+    };
+  });
+}
+
+// src/addons/face-detection/index.ts
+var FACE_LABEL = { id: "face", name: "Face" };
+var FACE_LABELS2 = [FACE_LABEL];
+var FACE_CLASS_MAP = { mapping: {}, preserveOriginal: true };
+var FaceDetectionAddon = class {
+  id = "face-detection";
+  slot = "cropper";
+  inputClasses = ["person"];
+  outputClasses = ["face"];
+  slotPriority = 0;
+  manifest = {
+    id: "face-detection",
+    name: "Face Detection",
+    version: "0.1.0",
+    description: "SCRFD-based face detector \u2014 crops face regions from person detections",
+    packageName: "@camstack/addon-vision",
+    slot: "cropper",
+    inputClasses: ["person"],
+    outputClasses: ["face"],
+    supportsCustomModels: false,
+    mayRequirePython: false,
+    defaultConfig: {
+      modelId: "scrfd-500m",
+      runtime: "auto",
+      backend: "cpu",
+      confidence: 0.5
+    }
+  };
+  engine;
+  modelEntry;
+  confidence = 0.5;
+  async initialize(ctx) {
+    const cfg = ctx.addonConfig;
+    const modelId = cfg["modelId"] ?? "scrfd-500m";
+    const runtime = cfg["runtime"] ?? "auto";
+    const backend = cfg["backend"] ?? "cpu";
+    this.confidence = cfg["confidence"] ?? 0.5;
+    const entry = FACE_DETECTION_MODELS.find((m) => m.id === modelId);
+    if (!entry) {
+      throw new Error(`FaceDetectionAddon: unknown modelId "${modelId}"`);
+    }
+    this.modelEntry = entry;
+    const resolved = await resolveEngine({
+      runtime,
+      backend,
+      modelEntry: entry,
+      modelsDir: ctx.locationPaths.models
+    });
+    this.engine = resolved.engine;
+  }
+  async crop(input) {
+    const start = Date.now();
+    const { width: inputW, height: inputH } = this.modelEntry.inputSize;
+    const targetSize = Math.max(inputW, inputH);
+    const personCrop = await cropRegion(input.frame.data, input.roi);
+    const lb = await letterbox(personCrop, targetSize);
+    const engineWithMulti = this.engine;
+    let outputs;
+    if (typeof engineWithMulti.runMultiOutput === "function") {
+      outputs = await engineWithMulti.runMultiOutput(lb.data, [1, 3, targetSize, targetSize]);
+    } else {
+      const single = await this.engine.run(lb.data, [1, 3, targetSize, targetSize]);
+      outputs = { output0: single };
+    }
+    const crops = scrfdPostprocess(
+      outputs,
+      this.confidence,
+      targetSize,
+      lb.originalWidth,
+      lb.originalHeight
+    );
+    return {
+      crops,
+      inferenceMs: Date.now() - start,
+      modelId: this.modelEntry.id
+    };
+  }
+  async shutdown() {
+    await this.engine?.dispose();
+  }
+  getConfigSchema() {
+    return {
+      sections: [
+        {
+          id: "model",
+          title: "Model",
+          columns: 1,
+          fields: [
+            {
+              key: "modelId",
+              label: "Model",
+              type: "model-selector",
+              catalog: [...FACE_DETECTION_MODELS],
+              allowCustom: false,
+              allowConversion: false,
+              acceptFormats: ["onnx", "coreml", "openvino"],
+              requiredMetadata: ["inputSize", "labels", "outputFormat"],
+              outputFormatHint: "ssd"
+            }
+          ]
+        },
+        {
+          id: "thresholds",
+          title: "Detection Thresholds",
+          columns: 1,
+          fields: [
+            {
+              key: "confidence",
+              label: "Confidence Threshold",
+              type: "slider",
+              min: 0.1,
+              max: 1,
+              step: 0.05,
+              default: 0.5
+            }
+          ]
+        }
+      ]
+    };
+  }
+  getClassMap() {
+    return FACE_CLASS_MAP;
+  }
+  getModelCatalog() {
+    return [...FACE_DETECTION_MODELS];
+  }
+  getAvailableModels() {
+    return [];
+  }
+  getActiveLabels() {
+    return FACE_LABELS2;
+  }
+  async probe() {
+    return {
+      available: true,
+      runtime: this.engine?.runtime ?? "onnx",
+      device: this.engine?.device ?? "cpu",
+      capabilities: ["fp32"]
+    };
+  }
+};
+
+export {
+  scrfdPostprocess,
+  FACE_DETECTION_MODELS,
+  FaceDetectionAddon
+};
+//# sourceMappingURL=chunk-MEVASN3P.mjs.map

package/dist/chunk-MEVASN3P.mjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/catalogs/face-detection-models.ts","../src/shared/postprocess/scrfd.ts","../src/addons/face-detection/index.ts"],"sourcesContent":["import type { ModelCatalogEntry, LabelDefinition } from '@camstack/types'\nimport { hfModelUrl } from '@camstack/types'\n\nconst HF_REPO = 'camstack/camstack-models'\n\nconst FACE_LABELS: readonly LabelDefinition[] = [\n  { id: 'face', name: 'Face' },\n] as const\n\nexport const FACE_DETECTION_MODELS: readonly ModelCatalogEntry[] = [\n  {\n    id: 'scrfd-500m',\n    name: 'SCRFD 500M',\n    description: 'SCRFD 500M — ultra-lightweight face detector',\n    inputSize: { width: 640, height: 640 },\n    labels: FACE_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/onnx/camstack-scrfd-500m.onnx'),\n        sizeMB: 2.2,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/coreml/camstack-scrfd-500m.mlpackage'),\n        sizeMB: 1.2,\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/openvino/camstack-scrfd-500m.xml'),\n        sizeMB: 1.3,\n      },\n    },\n  },\n  {\n    id: 'scrfd-2.5g',\n    name: 'SCRFD 2.5G',\n    description: 'SCRFD 2.5G — balanced face detection model',\n    inputSize: { width: 640, height: 640 },\n    labels: FACE_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/onnx/camstack-scrfd-2.5g.onnx'),\n        sizeMB: 3.1,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/coreml/camstack-scrfd-2.5g.mlpackage'),\n        sizeMB: 1.7,\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/openvino/camstack-scrfd-2.5g.xml'),\n        sizeMB: 1.8,\n      },\n    },\n  },\n  {\n    id: 'scrfd-10g',\n    name: 'SCRFD 10G',\n    description: 'SCRFD 10G — high-accuracy face detector',\n    inputSize: { width: 640, height: 640 },\n    labels: FACE_LABELS,\n    formats: {\n      onnx: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/onnx/camstack-scrfd-10g.onnx'),\n        sizeMB: 16,\n      },\n      coreml: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/coreml/camstack-scrfd-10g.mlpackage'),\n        sizeMB: 8.2,\n      },\n      openvino: {\n        url: hfModelUrl(HF_REPO, 'faceDetection/scrfd/openvino/camstack-scrfd-10g.xml'),\n        sizeMB: 8.3,\n      },\n    },\n  },\n] as const\n","import type { SpatialDetection, BoundingBox, Landmark } from '@camstack/types'\nimport { iou, nms } from './yolo.js'\n\nexport interface ScrfdRawOutputs {\n  readonly [key: string]: Float32Array\n}\n\nconst STRIDES = [8, 16, 32] as const\nconst NUM_ANCHORS_PER_STRIDE = 2\n\n/** Generate anchor points for a given stride and input size */\nfunction generateAnchors(stride: number, inputSize: number): Array<{ cx: number; cy: number }> {\n  const featureSize = Math.ceil(inputSize / stride)\n  const anchors: Array<{ cx: number; cy: number }> = []\n  for (let y = 0; y < featureSize; y++) {\n    for (let x = 0; x < featureSize; x++) {\n      for (let k = 0; k < NUM_ANCHORS_PER_STRIDE; k++) {\n        anchors.push({\n          cx: (x + 0.5) * stride,\n          cy: (y + 0.5) * stride,\n        })\n      }\n    }\n  }\n  return anchors\n}\n\nexport function scrfdPostprocess(\n  outputs: ScrfdRawOutputs,\n  confidence: number,\n  inputSize: number,\n  originalWidth: number,\n  originalHeight: number,\n): SpatialDetection[] {\n  // Scale factor from letterbox (assume square crop, so same scale both axes)\n  const scaleX = originalWidth / inputSize\n  const scaleY = originalHeight / inputSize\n\n  interface Candidate {\n    readonly bbox: BoundingBox\n    readonly score: number\n    readonly landmarks?: readonly Landmark[]\n  }\n\n  const candidates: Candidate[] = []\n\n  for (const stride of STRIDES) {\n    const scoreKey = Object.keys(outputs).find((k) => k.includes(`score_${stride}`) || k.includes(`_${stride}_score`))\n    const bboxKey = Object.keys(outputs).find((k) => k.includes(`bbox_${stride}`) || k.includes(`_${stride}_bbox`))\n    const kpsKey = Object.keys(outputs).find((k) => k.includes(`kps_${stride}`) || k.includes(`_${stride}_kps`))\n\n    if (!scoreKey || !bboxKey) continue\n\n    const scores = outputs[scoreKey]!\n    const bboxes = outputs[bboxKey]!\n    const kps = kpsKey ? outputs[kpsKey] : undefined\n    const anchors = generateAnchors(stride, inputSize)\n\n    const n = anchors.length\n\n    for (let i = 0; i < n; i++) {\n      const score = scores[i]!\n      if (score < confidence) continue\n\n      const anchor = anchors[i]!\n\n      // Bboxes are relative to the anchor center in stride units, scaled by stride\n      const x1 = anchor.cx - bboxes[i * 4]! * stride\n      const y1 = anchor.cy - bboxes[i * 4 + 1]! * stride\n      const x2 = anchor.cx + bboxes[i * 4 + 2]! * stride\n      const y2 = anchor.cy + bboxes[i * 4 + 3]! * stride\n\n      const bbox: BoundingBox = {\n        x: x1 * scaleX,\n        y: y1 * scaleY,\n        w: (x2 - x1) * scaleX,\n        h: (y2 - y1) * scaleY,\n      }\n\n      let landmarks: readonly Landmark[] | undefined\n      if (kps) {\n        const pts: Landmark[] = []\n        for (let p = 0; p < 5; p++) {\n          pts.push({\n            x: (anchor.cx + kps[i * 10 + p * 2]! * stride) * scaleX,\n            y: (anchor.cy + kps[i * 10 + p * 2 + 1]! * stride) * scaleY,\n          })\n        }\n        landmarks = pts\n      }\n\n      candidates.push({ bbox, score, landmarks })\n    }\n  }\n\n  if (candidates.length === 0) return []\n\n  const keptIndices = nms(candidates, 0.45)\n\n  return keptIndices.map((idx) => {\n    const { bbox, score, landmarks } = candidates[idx]!\n    return {\n      class: 'face',\n      originalClass: 'face',\n      score,\n      bbox,\n      ...(landmarks ? { landmarks } : {}),\n    } satisfies SpatialDetection\n  })\n}\n","import type {\n  ICropperProvider,\n  IDetectionAddon,\n  AddonManifest,\n  AddonContext,\n  CropInput,\n  CropperOutput,\n  ConfigUISchema,\n  ClassMapDefinition,\n  ProbeResult,\n  ModelCatalogEntry,\n  DetectionModel,\n  LabelDefinition,\n  IInferenceEngine,\n} from '@camstack/types'\nimport { FACE_DETECTION_MODELS } from '../../catalogs/face-detection-models.js'\nimport { cropRegion, letterbox } from '../../shared/image-utils.js'\nimport { scrfdPostprocess } from '../../shared/postprocess/scrfd.js'\nimport { resolveEngine } from '../../shared/engine-resolver.js'\n\nconst FACE_LABEL: LabelDefinition = { id: 'face', name: 'Face' }\nconst FACE_LABELS: readonly LabelDefinition[] = [FACE_LABEL]\nconst FACE_CLASS_MAP: ClassMapDefinition = { mapping: {}, preserveOriginal: true }\n\nexport default class FaceDetectionAddon implements ICropperProvider, IDetectionAddon {\n  readonly id = 'face-detection'\n  readonly slot = 'cropper' as const\n  readonly inputClasses = ['person'] as const\n  readonly outputClasses = ['face'] as const\n  readonly slotPriority = 0\n  readonly manifest: AddonManifest = {\n    id: 'face-detection',\n    name: 'Face Detection',\n    version: '0.1.0',\n    description: 'SCRFD-based face detector — crops face regions from person detections',\n    packageName: '@camstack/addon-vision',\n    slot: 'cropper',\n    inputClasses: ['person'],\n    outputClasses: ['face'],\n    supportsCustomModels: false,\n    mayRequirePython: false,\n    defaultConfig: {\n      modelId: 'scrfd-500m',\n      runtime: 'auto',\n      backend: 'cpu',\n      confidence: 0.5,\n    },\n  }\n\n  private engine!: IInferenceEngine\n  private modelEntry!: ModelCatalogEntry\n  private confidence = 0.5\n\n  async initialize(ctx: AddonContext): Promise<void> {\n    const cfg = ctx.addonConfig\n    const modelId = (cfg['modelId'] as string | undefined) ?? 'scrfd-500m'\n    const runtime = (cfg['runtime'] as string | undefined) ?? 'auto'\n    const backend = (cfg['backend'] as string | undefined) ?? 'cpu'\n    this.confidence = (cfg['confidence'] as number | undefined) ?? 0.5\n\n    const entry = FACE_DETECTION_MODELS.find((m) => m.id === modelId)\n    if (!entry) {\n      throw new Error(`FaceDetectionAddon: unknown modelId \"${modelId}\"`)\n    }\n    this.modelEntry = entry\n\n    const resolved = await resolveEngine({\n      runtime: runtime as 'auto',\n      backend,\n      modelEntry: entry,\n      modelsDir: ctx.locationPaths.models,\n    })\n    this.engine = resolved.engine\n  }\n\n  async crop(input: CropInput): Promise<CropperOutput> {\n    const start = Date.now()\n    const { width: inputW, height: inputH } = this.modelEntry.inputSize\n    const targetSize = Math.max(inputW, inputH)\n\n    // Crop the person region from the full frame\n    const personCrop = await cropRegion(input.frame.data, input.roi)\n\n    // Letterbox resize to model input size\n    const lb = await letterbox(personCrop, targetSize)\n\n    // SCRFD has multiple output tensors\n    const engineWithMulti = this.engine as IInferenceEngine & {\n      runMultiOutput?: (input: Float32Array, shape: readonly number[]) => Promise<Record<string, Float32Array>>\n    }\n\n    let outputs: Record<string, Float32Array>\n    if (typeof engineWithMulti.runMultiOutput === 'function') {\n      outputs = await engineWithMulti.runMultiOutput(lb.data, [1, 3, targetSize, targetSize])\n    } else {\n      // Fallback: wrap single output\n      const single = await this.engine.run(lb.data, [1, 3, targetSize, targetSize])\n      outputs = { output0: single }\n    }\n\n    const crops = scrfdPostprocess(\n      outputs,\n      this.confidence,\n      targetSize,\n      lb.originalWidth,\n      lb.originalHeight,\n    )\n\n    return {\n      crops,\n      inferenceMs: Date.now() - start,\n      modelId: this.modelEntry.id,\n    }\n  }\n\n  async shutdown(): Promise<void> {\n    await this.engine?.dispose()\n  }\n\n  getConfigSchema(): ConfigUISchema {\n    return {\n      sections: [\n        {\n          id: 'model',\n          title: 'Model',\n          columns: 1,\n          fields: [\n            {\n              key: 'modelId',\n              label: 'Model',\n              type: 'model-selector',\n              catalog: [...FACE_DETECTION_MODELS],\n              allowCustom: false,\n              allowConversion: false,\n              acceptFormats: ['onnx', 'coreml', 'openvino'],\n              requiredMetadata: ['inputSize', 'labels', 'outputFormat'],\n              outputFormatHint: 'ssd',\n            },\n          ],\n        },\n        {\n          id: 'thresholds',\n          title: 'Detection Thresholds',\n          columns: 1,\n          fields: [\n            {\n              key: 'confidence',\n              label: 'Confidence Threshold',\n              type: 'slider',\n              min: 0.1,\n              max: 1.0,\n              step: 0.05,\n              default: 0.5,\n            },\n          ],\n        },\n      ],\n    }\n  }\n\n  getClassMap(): ClassMapDefinition {\n    return FACE_CLASS_MAP\n  }\n\n  getModelCatalog(): ModelCatalogEntry[] {\n    return [...FACE_DETECTION_MODELS]\n  }\n\n  getAvailableModels(): DetectionModel[] {\n    return []\n  }\n\n  getActiveLabels(): readonly LabelDefinition[] {\n    return FACE_LABELS\n  }\n\n  async probe(): Promise<ProbeResult> {\n    return {\n      available: true,\n      runtime: this.engine?.runtime ?? 'onnx',\n      device: this.engine?.device ?? 'cpu',\n      capabilities: ['fp32'],\n    }\n  }\n}\n"],"mappings":";;;;;;;;;;;;AACA,SAAS,kBAAkB;AAE3B,IAAM,UAAU;AAEhB,IAAM,cAA0C;AAAA,EAC9C,EAAE,IAAI,QAAQ,MAAM,OAAO;AAC7B;AAEO,IAAM,wBAAsD;AAAA,EACjE;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,aAAa;AAAA,IACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAI;AAAA,IACrC,QAAQ;AAAA,IACR,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,KAAK,WAAW,SAAS,mDAAmD;AAAA,QAC5E,QAAQ;AAAA,MACV;AAAA,MACA,QAAQ;AAAA,QACN,KAAK,WAAW,SAAS,0DAA0D;AAAA,QACnF,QAAQ;AAAA,MACV;AAAA,MACA,UAAU;AAAA,QACR,KAAK,WAAW,SAAS,sDAAsD;AAAA,QAC/E,QAAQ;AAAA,MACV;AAAA,IACF;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,aAAa;AAAA,IACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAI;AAAA,IACrC,QAAQ;AAAA,IACR,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,KAAK,WAAW,SAAS,mDAAmD;AAAA,QAC5E,QAAQ;AAAA,MACV;AAAA,MACA,QAAQ;AAAA,QACN,KAAK,WAAW,SAAS,0DAA0D;AAAA,QACnF,QAAQ;AAAA,MACV;AAAA,MACA,UAAU;AAAA,QACR,KAAK,WAAW,SAAS,sDAAsD;AAAA,QAC/E,QAAQ;AAAA,MACV;AAAA,IACF;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,aAAa;AAAA,IACb,WAAW,EAAE,OAAO,KAAK,QAAQ,IAAI;AAAA,IACrC,QAAQ;AAAA,IACR,SAAS;AAAA,MACP,MAAM;AAAA,QACJ,KAAK,WAAW,SAAS,kDAAkD;AAAA,QAC3E,QAAQ;AAAA,MACV;AAAA,MACA,QAAQ;AAAA,QACN,KAAK,WAAW,SAAS,yDAAyD;AAAA,QAClF,QAAQ;AAAA,MACV;AAAA,MACA,UAAU;AAAA,QACR,KAAK,WAAW,SAAS,qDAAqD;AAAA,QAC9E,QAAQ;AAAA,MACV;AAAA,IACF;AAAA,EACF;AACF;;;AClEA,IAAM,UAAU,CAAC,GAAG,IAAI,EAAE;AAC1B,IAAM,yBAAyB;AAG/B,SAAS,gBAAgB,QAAgB,WAAsD;AAC7F,QAAM,cAAc,KAAK,KAAK,YAAY,MAAM;AAChD,QAAM,UAA6C,CAAC;AACpD,WAAS,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,aAAS,IAAI,GAAG,IAAI,aAAa,KAAK;AACpC,eAAS,IAAI,GAAG,IAAI,wBAAwB,KAAK;AAC/C,gBAAQ,KAAK;AAAA,UACX,KAAK,IAAI,OAAO;AAAA,UAChB,KAAK,IAAI,OAAO;AAAA,QAClB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,iBACd,SACA,YACA,WACA,eACA,gBACoB;AAEpB,QAAM,SAAS,gBAAgB;AAC/B,QAAM,SAAS,iBAAiB;AAQhC,QAAM,aAA0B,CAAC;AAEjC,aAAW,UAAU,SAAS;AAC5B,UAAM,WAAW,OAAO,KAAK,OAAO,EAAE,KAAK,CAAC,MAAM,EAAE,SAAS,SAAS,MAAM,EAAE,KAAK,EAAE,SAAS,IAAI,MAAM,QAAQ,CAAC;AACjH,UAAM,UAAU,OAAO,KAAK,OAAO,EAAE,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,MAAM,EAAE,KAAK,EAAE,SAAS,IAAI,MAAM,OAAO,CAAC;AAC9G,UAAM,SAAS,OAAO,KAAK,OAAO,EAAE,KAAK,CAAC,MAAM,EAAE,SAAS,OAAO,MAAM,EAAE,KAAK,EAAE,SAAS,IAAI,MAAM,MAAM,CAAC;AAE3G,QAAI,CAAC,YAAY,CAAC,QAAS;AAE3B,UAAM,SAAS,QAAQ,QAAQ;AAC/B,UAAM,SAAS,QAAQ,OAAO;AAC9B,UAAM,MAAM,SAAS,QAAQ,MAAM,IAAI;AACvC,UAAM,UAAU,gBAAgB,QAAQ,SAAS;AAEjD,UAAM,IAAI,QAAQ;AAElB,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,QAAQ,OAAO,CAAC;AACtB,UAAI,QAAQ,WAAY;AAExB,YAAM,SAAS,QAAQ,CAAC;AAGxB,YAAM,KAAK,OAAO,KAAK,OAAO,IAAI,CAAC,IAAK;AACxC,YAAM,KAAK,OAAO,KAAK,OAAO,IAAI,IAAI,CAAC,IAAK;AAC5C,YAAM,KAAK,OAAO,KAAK,OAAO,IAAI,IAAI,CAAC,IAAK;AAC5C,YAAM,KAAK,OAAO,KAAK,OAAO,IAAI,IAAI,CAAC,IAAK;AAE5C,YAAM,OAAoB;AAAA,QACxB,GAAG,KAAK;AAAA,QACR,GAAG,KAAK;AAAA,QACR,IAAI,KAAK,MAAM;AAAA,QACf,IAAI,KAAK,MAAM;AAAA,MACjB;AAEA,UAAI;AACJ,UAAI,KAAK;AACP,cAAM,MAAkB,CAAC;AACzB,iBAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,cAAI,KAAK;AAAA,YACP,IAAI,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI,CAAC,IAAK,UAAU;AAAA,YACjD,IAAI,OAAO,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,IAAK,UAAU;AAAA,UACvD,CAAC;AAAA,QACH;AACA,oBAAY;AAAA,MACd;AAEA,iBAAW,KAAK,EAAE,MAAM,OAAO,UAAU,CAAC;AAAA,IAC5C;AAAA,EACF;AAEA,MAAI,WAAW,WAAW,EAAG,QAAO,CAAC;AAErC,QAAM,cAAc,IAAI,YAAY,IAAI;AAExC,SAAO,YAAY,IAAI,CAAC,QAAQ;AAC9B,UAAM,EAAE,MAAM,OAAO,UAAU,IAAI,WAAW,GAAG;AACjD,WAAO;AAAA,MACL,OAAO;AAAA,MACP,eAAe;AAAA,MACf;AAAA,MACA;AAAA,MACA,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC;AAAA,IACnC;AAAA,EACF,CAAC;AACH;;;ACzFA,IAAM,aAA8B,EAAE,IAAI,QAAQ,MAAM,OAAO;AAC/D,IAAMA,eAA0C,CAAC,UAAU;AAC3D,IAAM,iBAAqC,EAAE,SAAS,CAAC,GAAG,kBAAkB,KAAK;AAEjF,IAAqB,qBAArB,MAAqF;AAAA,EAC1E,KAAK;AAAA,EACL,OAAO;AAAA,EACP,eAAe,CAAC,QAAQ;AAAA,EACxB,gBAAgB,CAAC,MAAM;AAAA,EACvB,eAAe;AAAA,EACf,WAA0B;AAAA,IACjC,IAAI;AAAA,IACJ,MAAM;AAAA,IACN,SAAS;AAAA,IACT,aAAa;AAAA,IACb,aAAa;AAAA,IACb,MAAM;AAAA,IACN,cAAc,CAAC,QAAQ;AAAA,IACvB,eAAe,CAAC,MAAM;AAAA,IACtB,sBAAsB;AAAA,IACtB,kBAAkB;AAAA,IAClB,eAAe;AAAA,MACb,SAAS;AAAA,MACT,SAAS;AAAA,MACT,SAAS;AAAA,MACT,YAAY;AAAA,IACd;AAAA,EACF;AAAA,EAEQ;AAAA,EACA;AAAA,EACA,aAAa;AAAA,EAErB,MAAM,WAAW,KAAkC;AACjD,UAAM,MAAM,IAAI;AAChB,UAAM,UAAW,IAAI,SAAS,KAA4B;AAC1D,UAAM,UAAW,IAAI,SAAS,KAA4B;AAC1D,UAAM,UAAW,IAAI,SAAS,KAA4B;AAC1D,SAAK,aAAc,IAAI,YAAY,KAA4B;AAE/D,UAAM,QAAQ,sBAAsB,KAAK,CAAC,MAAM,EAAE,OAAO,OAAO;AAChE,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,MAAM,wCAAwC,OAAO,GAAG;AAAA,IACpE;AACA,SAAK,aAAa;AAElB,UAAM,WAAW,MAAM,cAAc;AAAA,MACnC;AAAA,MACA;AAAA,MACA,YAAY;AAAA,MACZ,WAAW,IAAI,cAAc;AAAA,IAC/B,CAAC;AACD,SAAK,SAAS,SAAS;AAAA,EACzB;AAAA,EAEA,MAAM,KAAK,OAA0C;AACnD,UAAM,QAAQ,KAAK,IAAI;AACvB,UAAM,EAAE,OAAO,QAAQ,QAAQ,OAAO,IAAI,KAAK,WAAW;AAC1D,UAAM,aAAa,KAAK,IAAI,QAAQ,MAAM;AAG1C,UAAM,aAAa,MAAM,WAAW,MAAM,MAAM,MAAM,MAAM,GAAG;AAG/D,UAAM,KAAK,MAAM,UAAU,YAAY,UAAU;AAGjD,UAAM,kBAAkB,KAAK;AAI7B,QAAI;AACJ,QAAI,OAAO,gBAAgB,mBAAmB,YAAY;AACxD,gBAAU,MAAM,gBAAgB,eAAe,GAAG,MAAM,CAAC,GAAG,GAAG,YAAY,UAAU,CAAC;AAAA,IACxF,OAAO;AAEL,YAAM,SAAS,MAAM,KAAK,OAAO,IAAI,GAAG,MAAM,CAAC,GAAG,GAAG,YAAY,UAAU,CAAC;AAC5E,gBAAU,EAAE,SAAS,OAAO;AAAA,IAC9B;AAEA,UAAM,QAAQ;AAAA,MACZ;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA,GAAG;AAAA,MACH,GAAG;AAAA,IACL;AAEA,WAAO;AAAA,MACL;AAAA,MACA,aAAa,KAAK,IAAI,IAAI;AAAA,MAC1B,SAAS,KAAK,WAAW;AAAA,IAC3B;AAAA,EACF;AAAA,EAEA,MAAM,WAA0B;AAC9B,UAAM,KAAK,QAAQ,QAAQ;AAAA,EAC7B;AAAA,EAEA,kBAAkC;AAChC,WAAO;AAAA,MACL,UAAU;AAAA,QACR;AAAA,UACE,IAAI;AAAA,UACJ,OAAO;AAAA,UACP,SAAS;AAAA,UACT,QAAQ;AAAA,YACN;AAAA,cACE,KAAK;AAAA,cACL,OAAO;AAAA,cACP,MAAM;AAAA,cACN,SAAS,CAAC,GAAG,qBAAqB;AAAA,cAClC,aAAa;AAAA,cACb,iBAAiB;AAAA,cACjB,eAAe,CAAC,QAAQ,UAAU,UAAU;AAAA,cAC5C,kBAAkB,CAAC,aAAa,UAAU,cAAc;AAAA,cACxD,kBAAkB;AAAA,YACpB;AAAA,UACF;AAAA,QACF;AAAA,QACA;AAAA,UACE,IAAI;AAAA,UACJ,OAAO;AAAA,UACP,SAAS;AAAA,UACT,QAAQ;AAAA,YACN;AAAA,cACE,KAAK;AAAA,cACL,OAAO;AAAA,cACP,MAAM;AAAA,cACN,KAAK;AAAA,cACL,KAAK;AAAA,cACL,MAAM;AAAA,cACN,SAAS;AAAA,YACX;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA,EAEA,cAAkC;AAChC,WAAO;AAAA,EACT;AAAA,EAEA,kBAAuC;AACrC,WAAO,CAAC,GAAG,qBAAqB;AAAA,EAClC;AAAA,EAEA,qBAAuC;AACrC,WAAO,CAAC;AAAA,EACV;AAAA,EAEA,kBAA8C;AAC5C,WAAOA;AAAA,EACT;AAAA,EAEA,MAAM,QAA8B;AAClC,WAAO;AAAA,MACL,WAAW;AAAA,MACX,SAAS,KAAK,QAAQ,WAAW;AAAA,MACjC,QAAQ,KAAK,QAAQ,UAAU;AAAA,MAC/B,cAAc,CAAC,MAAM;AAAA,IACvB;AAAA,EACF;AACF;","names":["FACE_LABELS"]}

package/dist/chunk-PDSHDDPV.mjs

@@ -0,0 +1,255 @@
+import {
+  cropRegion,
+  resizeAndNormalize
+} from "./chunk-6OR5TE7A.mjs";
+import {
+  resolveEngine
+} from "./chunk-LPI42WL6.mjs";
+
+// src/catalogs/plate-recognition-models.ts
+import { hfModelUrl } from "@camstack/types";
+var HF_REPO = "camstack/camstack-models";
+var PLATE_TEXT_LABELS = [
+  { id: "text", name: "Plate Text" }
+];
+var PLATE_RECOGNITION_MODELS = [
+  {
+    id: "paddleocr-latin",
+    name: "PaddleOCR Latin",
+    description: "PaddleOCR recognition model for Latin-script license plates",
+    inputSize: { width: 320, height: 48 },
+    labels: PLATE_TEXT_LABELS,
+    formats: {
+      onnx: {
+        url: hfModelUrl(HF_REPO, "plateRecognition/paddleocr/onnx/camstack-paddleocr-latin-rec.onnx"),
+        sizeMB: 7.5
+      },
+      openvino: {
+        url: hfModelUrl(HF_REPO, "plateRecognition/paddleocr/openvino/camstack-paddleocr-latin.xml"),
+        sizeMB: 4
+      }
+    }
+  },
+  {
+    id: "paddleocr-en",
+    name: "PaddleOCR English",
+    description: "PaddleOCR recognition model optimized for English license plates",
+    inputSize: { width: 320, height: 48 },
+    labels: PLATE_TEXT_LABELS,
+    formats: {
+      onnx: {
+        url: hfModelUrl(HF_REPO, "plateRecognition/paddleocr/onnx/camstack-paddleocr-en-rec.onnx"),
+        sizeMB: 7.5
+      },
+      openvino: {
+        url: hfModelUrl(HF_REPO, "plateRecognition/paddleocr/openvino/camstack-paddleocr-en.xml"),
+        sizeMB: 4
+      }
+    }
+  }
+];
+
+// src/shared/postprocess/paddleocr.ts
+function ctcDecode(output, seqLen, numChars, charset) {
+  let totalLogScore = 0;
+  const rawIndices = [];
+  for (let t = 0; t < seqLen; t++) {
+    const offset = t * numChars;
+    let bestIdx = 0;
+    let bestVal = output[offset];
+    for (let c = 1; c < numChars; c++) {
+      const val = output[offset + c];
+      if (val > bestVal) {
+        bestVal = val;
+        bestIdx = c;
+      }
+    }
+    rawIndices.push(bestIdx);
+    totalLogScore += bestVal;
+  }
+  const collapsed = [];
+  for (let i = 0; i < rawIndices.length; i++) {
+    const cur = rawIndices[i];
+    if (i === 0 || cur !== rawIndices[i - 1]) {
+      collapsed.push(cur);
+    }
+  }
+  const filtered = collapsed.filter((idx) => idx !== 0);
+  const text = filtered.map((idx) => charset[idx] ?? "").join("");
+  const confidence = seqLen > 0 ? totalLogScore / seqLen : 0;
+  return { text, confidence };
+}
+
+// src/addons/plate-recognition/index.ts
+import * as fs from "fs";
+import * as path from "path";
+var PLATE_TEXT_LABEL = { id: "plate-text", name: "Plate Text" };
+var PLATE_TEXT_LABELS2 = [PLATE_TEXT_LABEL];
+var PLATE_REC_CLASS_MAP = { mapping: {}, preserveOriginal: true };
+function loadCharset(modelsDir, modelId) {
+  const dictNames = [
+    `camstack-${modelId}-dict.txt`,
+    `camstack-paddleocr-latin-dict.txt`,
+    `camstack-paddleocr-en-dict.txt`
+  ];
+  for (const name of dictNames) {
+    const dictPath = path.join(modelsDir, name);
+    if (fs.existsSync(dictPath)) {
+      const lines = fs.readFileSync(dictPath, "utf-8").split("\n").filter((l) => l.length > 0);
+      return ["", ...lines, " "];
+    }
+  }
+  throw new Error(`PlateRecognitionAddon: dict.txt not found in ${modelsDir}`);
+}
+var CHARSET = [];
+var REQUIRED_STEPS = [
+  { slot: "cropper", outputClasses: ["plate"], description: "Requires a plate detector" }
+];
+var PlateRecognitionAddon = class {
+  id = "plate-recognition";
+  slot = "classifier";
+  inputClasses = ["plate"];
+  outputClasses = ["plate-text:*"];
+  slotPriority = 0;
+  requiredSteps = REQUIRED_STEPS;
+  manifest = {
+    id: "plate-recognition",
+    name: "License Plate Recognition (OCR)",
+    version: "0.1.0",
+    description: "PaddleOCR-based license plate text recognition",
+    packageName: "@camstack/addon-vision",
+    slot: "classifier",
+    inputClasses: ["plate"],
+    outputClasses: ["plate-text:*"],
+    requiredSteps: REQUIRED_STEPS,
+    supportsCustomModels: false,
+    mayRequirePython: false,
+    defaultConfig: {
+      modelId: "paddleocr-latin",
+      runtime: "auto",
+      backend: "cpu",
+      minConfidence: 0.5
+    }
+  };
+  engine;
+  modelEntry;
+  minConfidence = 0.5;
+  async initialize(ctx) {
+    const cfg = ctx.addonConfig;
+    const modelId = cfg["modelId"] ?? "paddleocr-latin";
+    const runtime = cfg["runtime"] ?? "auto";
+    const backend = cfg["backend"] ?? "cpu";
+    this.minConfidence = cfg["minConfidence"] ?? 0.5;
+    const entry = PLATE_RECOGNITION_MODELS.find((m) => m.id === modelId);
+    if (!entry) {
+      throw new Error(`PlateRecognitionAddon: unknown modelId "${modelId}"`);
+    }
+    this.modelEntry = entry;
+    CHARSET = loadCharset(ctx.locationPaths.models, modelId);
+    const resolved = await resolveEngine({
+      runtime,
+      backend,
+      modelEntry: entry,
+      modelsDir: ctx.locationPaths.models
+    });
+    this.engine = resolved.engine;
+  }
+  async classify(input) {
+    const start = Date.now();
+    const { width: inputW, height: inputH } = this.modelEntry.inputSize;
+    const plateCrop = await cropRegion(input.frame.data, input.roi);
+    const normalized = await resizeAndNormalize(plateCrop, inputW, inputH, "zero-one", "nchw");
+    const output = await this.engine.run(normalized, [1, 3, inputH, inputW]);
+    const numChars = CHARSET.length;
+    const seqLen = output.length / numChars;
+    const { text, confidence } = ctcDecode(output, seqLen, numChars, CHARSET);
+    if (confidence < this.minConfidence || text.trim().length === 0) {
+      return {
+        classifications: [],
+        inferenceMs: Date.now() - start,
+        modelId: this.modelEntry.id
+      };
+    }
+    return {
+      classifications: [
+        {
+          class: "plate-text",
+          score: confidence,
+          text
+        }
+      ],
+      inferenceMs: Date.now() - start,
+      modelId: this.modelEntry.id
+    };
+  }
+  async shutdown() {
+    await this.engine?.dispose();
+  }
+  getConfigSchema() {
+    return {
+      sections: [
+        {
+          id: "model",
+          title: "Model",
+          columns: 1,
+          fields: [
+            {
+              key: "modelId",
+              label: "Model",
+              type: "model-selector",
+              catalog: [...PLATE_RECOGNITION_MODELS],
+              allowCustom: false,
+              allowConversion: false,
+              acceptFormats: ["onnx", "openvino"],
+              requiredMetadata: ["inputSize", "labels", "outputFormat"],
+              outputFormatHint: "ocr"
+            }
+          ]
+        },
+        {
+          id: "thresholds",
+          title: "Recognition Settings",
+          columns: 1,
+          fields: [
+            {
+              key: "minConfidence",
+              label: "Minimum Confidence",
+              type: "slider",
+              min: 0.1,
+              max: 1,
+              step: 0.05,
+              default: 0.5
+            }
+          ]
+        }
+      ]
+    };
+  }
+  getClassMap() {
+    return PLATE_REC_CLASS_MAP;
+  }
+  getModelCatalog() {
+    return [...PLATE_RECOGNITION_MODELS];
+  }
+  getAvailableModels() {
+    return [];
+  }
+  getActiveLabels() {
+    return PLATE_TEXT_LABELS2;
+  }
+  async probe() {
+    return {
+      available: true,
+      runtime: this.engine?.runtime ?? "onnx",
+      device: this.engine?.device ?? "cpu",
+      capabilities: ["fp32"]
+    };
+  }
+};
+
+export {
+  ctcDecode,
+  PLATE_RECOGNITION_MODELS,
+  PlateRecognitionAddon
+};
+//# sourceMappingURL=chunk-PDSHDDPV.mjs.map