depth-estimation 0.1.0__tar.gz → 0.1.2__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (102) hide show
  1. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/LICENSE +0 -0
  2. {depth_estimation-0.1.0/src/depth_estimation.egg-info → depth_estimation-0.1.2}/PKG-INFO +141 -7
  3. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/README.md +133 -3
  4. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/pyproject.toml +35 -4
  5. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/__init__.py +91 -3
  6. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/cli.py +197 -42
  7. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/configuration_utils.py +2 -0
  8. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/__init__.py +4 -7
  9. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/base_dataset.py +1 -1
  10. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/diode.py +2 -4
  11. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/folder.py +5 -3
  12. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/hub.py +37 -3
  13. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/kitti.py +5 -8
  14. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/nyu_depth_v2.py +10 -14
  15. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/data/transforms.py +47 -30
  16. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/evaluation/__init__.py +0 -0
  17. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/evaluation/evaluate.py +48 -33
  18. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/evaluation/metrics.py +71 -46
  19. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/evaluation/profiling.py +15 -13
  20. depth_estimation-0.1.2/src/depth_estimation/export.py +301 -0
  21. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/losses.py +6 -2
  22. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/modeling_utils.py +114 -4
  23. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/__init__.py +0 -0
  24. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/auto/__init__.py +0 -0
  25. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/auto/modeling_auto.py +0 -0
  26. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/auto/processing_auto.py +2 -1
  27. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v1/__init__.py +2 -0
  28. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v1/configuration_depth_anything_v1.py +1 -1
  29. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v1/modeling_depth_anything_v1.py +173 -43
  30. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v2/__init__.py +2 -0
  31. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v2/configuration_depth_anything_v2.py +4 -2
  32. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v2/modeling_depth_anything_v2.py +397 -126
  33. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v3/__init__.py +4 -0
  34. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v3/configuration_depth_anything_v3.py +6 -6
  35. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_anything_v3/modeling_depth_anything_v3.py +257 -77
  36. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_pro/__init__.py +2 -0
  37. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_pro/configuration_depth_pro.py +0 -0
  38. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/depth_pro/modeling_depth_pro.py +79 -27
  39. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/marigold_dc/__init__.py +2 -0
  40. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/marigold_dc/configuration_marigold_dc.py +0 -0
  41. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/marigold_dc/modeling_marigold_dc.py +68 -30
  42. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/midas/__init__.py +2 -0
  43. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/midas/configuration_midas.py +14 -0
  44. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/midas/modeling_midas.py +3 -1
  45. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/moge/__init__.py +2 -0
  46. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/moge/configuration_moge.py +0 -0
  47. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/moge/modeling_moge.py +1136 -421
  48. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/omnivggt/__init__.py +2 -0
  49. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/omnivggt/configuration_omnivggt.py +0 -0
  50. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/omnivggt/modeling_omnivggt.py +861 -283
  51. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/pixel_perfect_depth/__init__.py +2 -0
  52. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/pixel_perfect_depth/configuration_ppd.py +0 -0
  53. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/pixel_perfect_depth/modeling_ppd.py +76 -34
  54. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/vggt/__init__.py +2 -0
  55. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/vggt/configuration_vggt.py +0 -0
  56. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/vggt/modeling_vggt.py +663 -223
  57. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/zoedepth/__init__.py +2 -0
  58. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/zoedepth/configuration_zoedepth.py +0 -0
  59. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/models/zoedepth/modeling_zoedepth.py +20 -6
  60. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/output.py +0 -0
  61. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/pipeline_utils.py +94 -3
  62. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/pipelines/__init__.py +0 -0
  63. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/pipelines/depth_estimation.py +0 -0
  64. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/processing_utils.py +24 -15
  65. depth_estimation-0.1.2/src/depth_estimation/pruning.py +216 -0
  66. depth_estimation-0.1.2/src/depth_estimation/quantization.py +284 -0
  67. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/registry.py +1 -1
  68. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/trainer.py +19 -11
  69. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation/training_args.py +13 -6
  70. depth_estimation-0.1.2/src/depth_estimation/video.py +263 -0
  71. depth_estimation-0.1.2/src/depth_estimation/viz/__init__.py +45 -0
  72. depth_estimation-0.1.2/src/depth_estimation/viz/animate.py +89 -0
  73. depth_estimation-0.1.2/src/depth_estimation/viz/composite.py +85 -0
  74. depth_estimation-0.1.2/src/depth_estimation/viz/display.py +108 -0
  75. depth_estimation-0.1.2/src/depth_estimation/viz/error.py +72 -0
  76. {depth_estimation-0.1.0 → depth_estimation-0.1.2/src/depth_estimation.egg-info}/PKG-INFO +141 -7
  77. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation.egg-info/SOURCES.txt +19 -1
  78. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation.egg-info/requires.txt +7 -1
  79. depth_estimation-0.1.2/tests/test_auto.py +174 -0
  80. depth_estimation-0.1.2/tests/test_cli.py +279 -0
  81. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_config.py +48 -26
  82. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_data.py +85 -4
  83. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_evaluation.py +15 -3
  84. depth_estimation-0.1.2/tests/test_export_onnx.py +402 -0
  85. depth_estimation-0.1.2/tests/test_inference_all_models.py +161 -0
  86. depth_estimation-0.1.2/tests/test_inference_image.py +72 -0
  87. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_losses.py +10 -7
  88. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_model.py +13 -4
  89. depth_estimation-0.1.2/tests/test_package.py +20 -0
  90. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_pipeline.py +6 -3
  91. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_processor.py +87 -1
  92. depth_estimation-0.1.2/tests/test_pruning.py +168 -0
  93. depth_estimation-0.1.2/tests/test_quantization.py +225 -0
  94. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_training.py +24 -11
  95. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/tests/test_transforms.py +22 -4
  96. depth_estimation-0.1.2/tests/test_video.py +249 -0
  97. depth_estimation-0.1.2/tests/test_viz.py +191 -0
  98. depth_estimation-0.1.0/tests/test_auto.py +0 -142
  99. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/setup.cfg +0 -0
  100. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation.egg-info/dependency_links.txt +0 -0
  101. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation.egg-info/entry_points.txt +0 -0
  102. {depth_estimation-0.1.0 → depth_estimation-0.1.2}/src/depth_estimation.egg-info/top_level.txt +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: depth_estimation
3
- Version: 0.1.0
3
+ Version: 0.1.2
4
4
  Summary: A Transformers-style Python library for monocular depth estimation — inference, evaluation, and fine-tuning
5
5
  Author: shriarul5273
6
6
  License-Expression: MIT
@@ -9,7 +9,6 @@ Project-URL: Repository, https://github.com/shriarul5273/depth_estimation
9
9
  Project-URL: Issues, https://github.com/shriarul5273/depth_estimation/issues
10
10
  Keywords: depth-estimation,monocular-depth,computer-vision,deep-learning,pytorch,fine-tuning,training
11
11
  Classifier: Programming Language :: Python :: 3
12
- Classifier: Programming Language :: Python :: 3.9
13
12
  Classifier: Programming Language :: Python :: 3.10
14
13
  Classifier: Programming Language :: Python :: 3.11
15
14
  Classifier: Programming Language :: Python :: 3.12
@@ -18,10 +17,10 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
18
17
  Classifier: Topic :: Scientific/Engineering :: Image Recognition
19
18
  Classifier: Intended Audience :: Developers
20
19
  Classifier: Intended Audience :: Science/Research
21
- Requires-Python: >=3.9
20
+ Requires-Python: >=3.10
22
21
  Description-Content-Type: text/markdown
23
22
  License-File: LICENSE
24
- Requires-Dist: torch==2.9.0
23
+ Requires-Dist: torch>=2.0
25
24
  Requires-Dist: torchvision>=0.15
26
25
  Requires-Dist: Pillow>=9.0
27
26
  Requires-Dist: numpy>=1.24
@@ -39,25 +38,33 @@ Requires-Dist: tqdm
39
38
  Provides-Extra: dev
40
39
  Requires-Dist: pytest>=7.0; extra == "dev"
41
40
  Requires-Dist: pytest-cov; extra == "dev"
41
+ Requires-Dist: onnx>=1.14; extra == "dev"
42
+ Requires-Dist: onnxruntime; extra == "dev"
43
+ Requires-Dist: mypy>=1.10; extra == "dev"
44
+ Provides-Extra: export
45
+ Requires-Dist: onnx>=1.14; extra == "export"
42
46
  Dynamic: license-file
43
47
 
44
48
  # depth_estimation
45
49
 
46
50
  <p align="center">
51
+ <a href="https://github.com/shriarul5273/depth_estimation/actions/workflows/test.yml"><img alt="CI" src="https://github.com/shriarul5273/depth_estimation/actions/workflows/test.yml/badge.svg"></a>
47
52
  <a href="https://github.com/shriarul5273/depth_estimation/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/github/license/shriarul5273/depth_estimation?color=blue"></a>
48
53
  <a href="https://pypi.org/project/depth-estimation/"><img alt="PyPI" src="https://img.shields.io/pypi/v/depth-estimation"></a>
49
54
  <a href="https://pypi.org/project/depth-estimation/"><img alt="Python" src="https://img.shields.io/pypi/pyversions/depth-estimation"></a>
55
+ <a href="https://pepy.tech/projects/depth-estimation"><img alt="PyPI Downloads" src="https://static.pepy.tech/personalized-badge/depth-estimation?period=total&units=INTERNATIONAL_SYSTEM&left_color=GREY&right_color=GREEN&left_text=downloads"></a>
56
+ <a href="https://huggingface.co/spaces/shriarul5273/Depth-Estimation-Compare-demo"><img alt="Demo" src="https://img.shields.io/badge/🤗%20Hugging%20Face-Demo-blue"></a>
50
57
  </p>
51
58
 
52
59
  <h3 align="center">A unified Python library for monocular depth estimation</h3>
53
60
 
54
- <h3 align="center">Inference · Fine-Tuning · Evaluation · Dataset Loading</h3>
61
+ <h3 align="center">Inference · Video & Streaming · Visualization · Fine-Tuning · Evaluation · Dataset Loading</h3>
55
62
 
56
63
  ---
57
64
 
58
65
  `depth_estimation` is the model-definition framework for depth estimation. It provides a single, consistent API across **12 model families and 28 variants** — so you can swap models, compare them, and fine-tune them without rewriting your pipeline.
59
66
 
60
- It covers the full workflow end-to-end: run inference with one line, evaluate on standard benchmarks, and fine-tune on custom depth data — all with the same library.
67
+ It covers the full workflow end-to-end: run inference with one line, stream depth from video, visualize results, evaluate on standard benchmarks, and fine-tune on custom depth data — all with the same library.
61
68
 
62
69
  ## Installation
63
70
 
@@ -165,6 +172,79 @@ depth-estimate predict "images/*.jpg" --model depth-anything-v2-vitb --output-di
165
172
 
166
173
  </details>
167
174
 
175
+ <details>
176
+ <summary><b>Video & Streaming</b> — frame-by-frame depth from video, webcam, or image sequences</summary>
177
+
178
+ ```python
179
+ from depth_estimation import pipeline
180
+
181
+ pipe = pipeline("depth-estimation", model="depth-anything-v2-vitb")
182
+
183
+ # Stream a video file — yields DepthOutput per frame
184
+ for result in pipe.stream("video.mp4", temporal_smoothing=0.5):
185
+ depth = result.depth # (H, W) float32
186
+ colored = result.colored_depth # (H, W, 3) uint8
187
+ print(result.metadata["frame_index"])
188
+
189
+ # Webcam stream
190
+ for result in pipe.stream(0): # device index
191
+ ...
192
+
193
+ # Frame glob (sorted alphabetically)
194
+ for result in pipe.stream("frames/*.png"):
195
+ ...
196
+
197
+ # Write output video to disk
198
+ pipe.process_video(
199
+ "input.mp4",
200
+ "output_depth.mp4",
201
+ colormap="inferno",
202
+ side_by_side=True, # RGB | depth composite
203
+ temporal_smoothing=0.5,
204
+ )
205
+ ```
206
+
207
+ ```bash
208
+ # CLI — video prediction
209
+ depth-estimate predict video.mp4 --model depth-anything-v2-vitb --output depth_video.mp4
210
+ ```
211
+
212
+ See [docs/video.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/video.md).
213
+
214
+ </details>
215
+
216
+ <details>
217
+ <summary><b>Visualization</b> — depth maps, comparisons, overlays, 3D animations, error maps</summary>
218
+
219
+ ```python
220
+ from depth_estimation.viz import (
221
+ show_depth, compare_depths, overlay_depth,
222
+ create_anaglyph, animate_3d, plot_error_map,
223
+ )
224
+
225
+ # Display a depth result
226
+ show_depth(result, colormap="Spectral_r", title="Depth Anything V2")
227
+
228
+ # Side-by-side comparison of multiple models
229
+ compare_depths([result_v2, result_pro], labels=["DA V2", "DepthPro"], save="compare.png")
230
+
231
+ # Blend depth over RGB image
232
+ overlay = overlay_depth(image, result.depth, alpha=0.5, colormap="inferno")
233
+
234
+ # Red-cyan anaglyph stereo image
235
+ anaglyph = create_anaglyph(image, result.depth, baseline=0.065)
236
+
237
+ # Rotating 3D surface animation
238
+ animate_3d(image, result.depth, "rotation.gif", frames=60)
239
+
240
+ # Per-pixel error heatmap (requires ground truth)
241
+ plot_error_map(pred_depth, gt_depth, metric="abs_rel", save="errors.png")
242
+ ```
243
+
244
+ See [docs/viz.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/viz.md).
245
+
246
+ </details>
247
+
168
248
  <details>
169
249
  <summary><b>Evaluation</b> — standard benchmarks, custom predictions</summary>
170
250
 
@@ -192,7 +272,8 @@ See [docs/evaluation.md](https://github.com/shriarul5273/depth_estimation/blob/m
192
272
  <summary><b>Fine-Tuning</b> — any trainable model, any depth dataset</summary>
193
273
 
194
274
  ```python
195
- from depth_estimation import DepthAnythingV2Model, DepthTrainer, DepthTrainingArguments, load_dataset
275
+ from depth_estimation import DepthTrainer, DepthTrainingArguments, load_dataset
276
+ from depth_estimation.models.depth_anything_v2 import DepthAnythingV2Model
196
277
  from depth_estimation.data.transforms import get_train_transforms, get_val_transforms
197
278
 
198
279
  model = DepthAnythingV2Model.from_pretrained("depth-anything-v2-vits", for_training=True)
@@ -224,6 +305,59 @@ See [docs/data.md](https://github.com/shriarul5273/depth_estimation/blob/main/do
224
305
 
225
306
  </details>
226
307
 
308
+ <details>
309
+ <summary><b>ONNX Export</b> — deploy outside PyTorch</summary>
310
+
311
+ ```python
312
+ from depth_estimation import AutoDepthModel, export_onnx
313
+
314
+ model = AutoDepthModel.from_pretrained("depth-anything-v2-vitb")
315
+ export_onnx(model, "depth_anything_v2_vitb.onnx", input_size=518, verify=True)
316
+ ```
317
+
318
+ ```bash
319
+ # CLI
320
+ depth-estimate export --model depth-anything-v2-vitb --output model.onnx --verify
321
+ ```
322
+
323
+ Verified working for `depth-anything-v1/v2/v3`, `depth-pro`, and `moge`. `zoedepth` and `marigold-dc` are **not exportable** (both wrap an opaque external pipeline) — `export_onnx()` raises a clear error immediately rather than a doomed attempt. Requires the optional `onnx` package: `pip install "depth-estimation[export]"`. GPU inference on the exported file works too via `onnxruntime-gpu`. See [docs/export.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/export.md) for the full supported-models table, known limitations, and GPU setup.
324
+
325
+ </details>
326
+
327
+ <details>
328
+ <summary><b>Pruning</b> — reduce effective parameter count</summary>
329
+
330
+ ```python
331
+ from depth_estimation import AutoDepthModel, prune_model, compute_sparsity
332
+
333
+ model = AutoDepthModel.from_pretrained("depth-anything-v2-vitb")
334
+ prune_model(model, amount=0.3) # zero out the smallest-magnitude 30% of weights
335
+ print(compute_sparsity(model)["overall"]) # ~0.3
336
+
337
+ model.export_onnx("pruned.onnx", verify=True) # pruned models export like any other
338
+ ```
339
+
340
+ Pure PyTorch (`torch.nn.utils.prune`) — no special hardware or SDK. See [docs/pruning.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/pruning.md) for prune-aware fine-tuning and what pruning does/doesn't buy you (it zeros weights, it doesn't shrink tensors).
341
+
342
+ </details>
343
+
344
+ <details>
345
+ <summary><b>Quantization</b> — float16/bfloat16/int8, plus ONNX int8/uint8</summary>
346
+
347
+ ```python
348
+ from depth_estimation import AutoDepthModel, quantize_onnx
349
+
350
+ model = AutoDepthModel.from_pretrained("depth-anything-v2-vitb")
351
+ model.quantize(dtype="float16") # in-place GPU precision cast
352
+
353
+ model.export_onnx("model.onnx")
354
+ quantize_onnx("model.onnx", "model_uint8.onnx") # default weight_type="uint8", verify=True
355
+ ```
356
+
357
+ `verify` defaults to `True` — quantization accuracy is **model-dependent, not reliably safe**: testing across the 28 registered variants found `uint8` produces badly wrong output for several real pretrained checkpoints, not just an edge case. See [docs/quantization.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/quantization.md) — including which formats are actually verified working (`int16`/`uint16` are **not**, despite being accepted by the underlying APIs, and `int8` needs `onnxruntime>=1.26.0` for models with `Conv2d` layers).
358
+
359
+ </details>
360
+
227
361
  ---
228
362
 
229
363
  ## Adding a New Model
@@ -1,20 +1,23 @@
1
1
  # depth_estimation
2
2
 
3
3
  <p align="center">
4
+ <a href="https://github.com/shriarul5273/depth_estimation/actions/workflows/test.yml"><img alt="CI" src="https://github.com/shriarul5273/depth_estimation/actions/workflows/test.yml/badge.svg"></a>
4
5
  <a href="https://github.com/shriarul5273/depth_estimation/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/github/license/shriarul5273/depth_estimation?color=blue"></a>
5
6
  <a href="https://pypi.org/project/depth-estimation/"><img alt="PyPI" src="https://img.shields.io/pypi/v/depth-estimation"></a>
6
7
  <a href="https://pypi.org/project/depth-estimation/"><img alt="Python" src="https://img.shields.io/pypi/pyversions/depth-estimation"></a>
8
+ <a href="https://pepy.tech/projects/depth-estimation"><img alt="PyPI Downloads" src="https://static.pepy.tech/personalized-badge/depth-estimation?period=total&units=INTERNATIONAL_SYSTEM&left_color=GREY&right_color=GREEN&left_text=downloads"></a>
9
+ <a href="https://huggingface.co/spaces/shriarul5273/Depth-Estimation-Compare-demo"><img alt="Demo" src="https://img.shields.io/badge/🤗%20Hugging%20Face-Demo-blue"></a>
7
10
  </p>
8
11
 
9
12
  <h3 align="center">A unified Python library for monocular depth estimation</h3>
10
13
 
11
- <h3 align="center">Inference · Fine-Tuning · Evaluation · Dataset Loading</h3>
14
+ <h3 align="center">Inference · Video & Streaming · Visualization · Fine-Tuning · Evaluation · Dataset Loading</h3>
12
15
 
13
16
  ---
14
17
 
15
18
  `depth_estimation` is the model-definition framework for depth estimation. It provides a single, consistent API across **12 model families and 28 variants** — so you can swap models, compare them, and fine-tune them without rewriting your pipeline.
16
19
 
17
- It covers the full workflow end-to-end: run inference with one line, evaluate on standard benchmarks, and fine-tune on custom depth data — all with the same library.
20
+ It covers the full workflow end-to-end: run inference with one line, stream depth from video, visualize results, evaluate on standard benchmarks, and fine-tune on custom depth data — all with the same library.
18
21
 
19
22
  ## Installation
20
23
 
@@ -122,6 +125,79 @@ depth-estimate predict "images/*.jpg" --model depth-anything-v2-vitb --output-di
122
125
 
123
126
  </details>
124
127
 
128
+ <details>
129
+ <summary><b>Video & Streaming</b> — frame-by-frame depth from video, webcam, or image sequences</summary>
130
+
131
+ ```python
132
+ from depth_estimation import pipeline
133
+
134
+ pipe = pipeline("depth-estimation", model="depth-anything-v2-vitb")
135
+
136
+ # Stream a video file — yields DepthOutput per frame
137
+ for result in pipe.stream("video.mp4", temporal_smoothing=0.5):
138
+ depth = result.depth # (H, W) float32
139
+ colored = result.colored_depth # (H, W, 3) uint8
140
+ print(result.metadata["frame_index"])
141
+
142
+ # Webcam stream
143
+ for result in pipe.stream(0): # device index
144
+ ...
145
+
146
+ # Frame glob (sorted alphabetically)
147
+ for result in pipe.stream("frames/*.png"):
148
+ ...
149
+
150
+ # Write output video to disk
151
+ pipe.process_video(
152
+ "input.mp4",
153
+ "output_depth.mp4",
154
+ colormap="inferno",
155
+ side_by_side=True, # RGB | depth composite
156
+ temporal_smoothing=0.5,
157
+ )
158
+ ```
159
+
160
+ ```bash
161
+ # CLI — video prediction
162
+ depth-estimate predict video.mp4 --model depth-anything-v2-vitb --output depth_video.mp4
163
+ ```
164
+
165
+ See [docs/video.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/video.md).
166
+
167
+ </details>
168
+
169
+ <details>
170
+ <summary><b>Visualization</b> — depth maps, comparisons, overlays, 3D animations, error maps</summary>
171
+
172
+ ```python
173
+ from depth_estimation.viz import (
174
+ show_depth, compare_depths, overlay_depth,
175
+ create_anaglyph, animate_3d, plot_error_map,
176
+ )
177
+
178
+ # Display a depth result
179
+ show_depth(result, colormap="Spectral_r", title="Depth Anything V2")
180
+
181
+ # Side-by-side comparison of multiple models
182
+ compare_depths([result_v2, result_pro], labels=["DA V2", "DepthPro"], save="compare.png")
183
+
184
+ # Blend depth over RGB image
185
+ overlay = overlay_depth(image, result.depth, alpha=0.5, colormap="inferno")
186
+
187
+ # Red-cyan anaglyph stereo image
188
+ anaglyph = create_anaglyph(image, result.depth, baseline=0.065)
189
+
190
+ # Rotating 3D surface animation
191
+ animate_3d(image, result.depth, "rotation.gif", frames=60)
192
+
193
+ # Per-pixel error heatmap (requires ground truth)
194
+ plot_error_map(pred_depth, gt_depth, metric="abs_rel", save="errors.png")
195
+ ```
196
+
197
+ See [docs/viz.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/viz.md).
198
+
199
+ </details>
200
+
125
201
  <details>
126
202
  <summary><b>Evaluation</b> — standard benchmarks, custom predictions</summary>
127
203
 
@@ -149,7 +225,8 @@ See [docs/evaluation.md](https://github.com/shriarul5273/depth_estimation/blob/m
149
225
  <summary><b>Fine-Tuning</b> — any trainable model, any depth dataset</summary>
150
226
 
151
227
  ```python
152
- from depth_estimation import DepthAnythingV2Model, DepthTrainer, DepthTrainingArguments, load_dataset
228
+ from depth_estimation import DepthTrainer, DepthTrainingArguments, load_dataset
229
+ from depth_estimation.models.depth_anything_v2 import DepthAnythingV2Model
153
230
  from depth_estimation.data.transforms import get_train_transforms, get_val_transforms
154
231
 
155
232
  model = DepthAnythingV2Model.from_pretrained("depth-anything-v2-vits", for_training=True)
@@ -181,6 +258,59 @@ See [docs/data.md](https://github.com/shriarul5273/depth_estimation/blob/main/do
181
258
 
182
259
  </details>
183
260
 
261
+ <details>
262
+ <summary><b>ONNX Export</b> — deploy outside PyTorch</summary>
263
+
264
+ ```python
265
+ from depth_estimation import AutoDepthModel, export_onnx
266
+
267
+ model = AutoDepthModel.from_pretrained("depth-anything-v2-vitb")
268
+ export_onnx(model, "depth_anything_v2_vitb.onnx", input_size=518, verify=True)
269
+ ```
270
+
271
+ ```bash
272
+ # CLI
273
+ depth-estimate export --model depth-anything-v2-vitb --output model.onnx --verify
274
+ ```
275
+
276
+ Verified working for `depth-anything-v1/v2/v3`, `depth-pro`, and `moge`. `zoedepth` and `marigold-dc` are **not exportable** (both wrap an opaque external pipeline) — `export_onnx()` raises a clear error immediately rather than a doomed attempt. Requires the optional `onnx` package: `pip install "depth-estimation[export]"`. GPU inference on the exported file works too via `onnxruntime-gpu`. See [docs/export.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/export.md) for the full supported-models table, known limitations, and GPU setup.
277
+
278
+ </details>
279
+
280
+ <details>
281
+ <summary><b>Pruning</b> — reduce effective parameter count</summary>
282
+
283
+ ```python
284
+ from depth_estimation import AutoDepthModel, prune_model, compute_sparsity
285
+
286
+ model = AutoDepthModel.from_pretrained("depth-anything-v2-vitb")
287
+ prune_model(model, amount=0.3) # zero out the smallest-magnitude 30% of weights
288
+ print(compute_sparsity(model)["overall"]) # ~0.3
289
+
290
+ model.export_onnx("pruned.onnx", verify=True) # pruned models export like any other
291
+ ```
292
+
293
+ Pure PyTorch (`torch.nn.utils.prune`) — no special hardware or SDK. See [docs/pruning.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/pruning.md) for prune-aware fine-tuning and what pruning does/doesn't buy you (it zeros weights, it doesn't shrink tensors).
294
+
295
+ </details>
296
+
297
+ <details>
298
+ <summary><b>Quantization</b> — float16/bfloat16/int8, plus ONNX int8/uint8</summary>
299
+
300
+ ```python
301
+ from depth_estimation import AutoDepthModel, quantize_onnx
302
+
303
+ model = AutoDepthModel.from_pretrained("depth-anything-v2-vitb")
304
+ model.quantize(dtype="float16") # in-place GPU precision cast
305
+
306
+ model.export_onnx("model.onnx")
307
+ quantize_onnx("model.onnx", "model_uint8.onnx") # default weight_type="uint8", verify=True
308
+ ```
309
+
310
+ `verify` defaults to `True` — quantization accuracy is **model-dependent, not reliably safe**: testing across the 28 registered variants found `uint8` produces badly wrong output for several real pretrained checkpoints, not just an edge case. See [docs/quantization.md](https://github.com/shriarul5273/depth_estimation/blob/main/docs/quantization.md) — including which formats are actually verified working (`int16`/`uint16` are **not**, despite being accepted by the underlying APIs, and `int8` needs `onnxruntime>=1.26.0` for models with `Conv2d` layers).
311
+
312
+ </details>
313
+
184
314
  ---
185
315
 
186
316
  ## Adding a New Model
@@ -4,18 +4,17 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "depth_estimation"
7
- version = "0.1.0"
7
+ version = "0.1.2"
8
8
  authors = [
9
9
  { name = "shriarul5273" },
10
10
  ]
11
11
  description = "A Transformers-style Python library for monocular depth estimation — inference, evaluation, and fine-tuning"
12
12
  readme = "README.md"
13
- requires-python = ">=3.9"
13
+ requires-python = ">=3.10"
14
14
  license = "MIT"
15
15
  license-files = ["LICENSE"]
16
16
  classifiers = [
17
17
  "Programming Language :: Python :: 3",
18
- "Programming Language :: Python :: 3.9",
19
18
  "Programming Language :: Python :: 3.10",
20
19
  "Programming Language :: Python :: 3.11",
21
20
  "Programming Language :: Python :: 3.12",
@@ -27,7 +26,7 @@ classifiers = [
27
26
  ]
28
27
  keywords = ["depth-estimation", "monocular-depth", "computer-vision", "deep-learning", "pytorch", "fine-tuning", "training"]
29
28
  dependencies = [
30
- "torch==2.9.0",
29
+ "torch>=2.0",
31
30
  "torchvision>=0.15",
32
31
  "Pillow>=9.0",
33
32
  "numpy>=1.24",
@@ -53,6 +52,12 @@ Issues = "https://github.com/shriarul5273/depth_estimation/issues"
53
52
  dev = [
54
53
  "pytest>=7.0",
55
54
  "pytest-cov",
55
+ "onnx>=1.14",
56
+ "onnxruntime",
57
+ "mypy>=1.10",
58
+ ]
59
+ export = [
60
+ "onnx>=1.14",
56
61
  ]
57
62
 
58
63
  [project.scripts]
@@ -60,3 +65,29 @@ depth-estimate = "depth_estimation.cli:main"
60
65
 
61
66
  [tool.setuptools.packages.find]
62
67
  where = ["src"]
68
+
69
+ [tool.pytest.ini_options]
70
+ markers = [
71
+ "slow: heavy or network-dependent tests, excluded by default (run with -m slow)",
72
+ ]
73
+
74
+ [tool.ruff.lint]
75
+ # Keep Ruff's original default rule set explicit. Ruff 0.16 broadened its
76
+ # implicit defaults, which made unchanged code fail CI when the unpinned
77
+ # runner upgraded from 0.15.x.
78
+ select = ["E4", "E7", "E9", "F"]
79
+
80
+ [tool.ruff.lint.per-file-ignores]
81
+ # Model submodules are imported for their self-registration side effect
82
+ # (they call MODEL_REGISTRY.register(...) at import time) — not because
83
+ # their names are used directly here.
84
+ "src/depth_estimation/__init__.py" = ["F401"]
85
+
86
+ [tool.mypy]
87
+ python_version = "3.10"
88
+ ignore_missing_imports = true
89
+ # The codebase currently has ~290 pre-existing type errors (mostly missing
90
+ # annotations in older modules, not real bugs — verified by spot-checking
91
+ # several by hand). Fixing all of them is a separate, large effort; CI runs
92
+ # mypy in advisory (non-blocking) mode for now — see .github/workflows/test.yml.
93
+ # New code should still aim to be clean under this config.
@@ -27,63 +27,130 @@ from .models import omnivggt
27
27
  from .models import vggt
28
28
 
29
29
 
30
- def load_dataset(name, split="train", root=None, download=True, transform=None, **kwargs):
30
+ def load_dataset(
31
+ name, split="train", root=None, download=True, transform=None, **kwargs
32
+ ):
31
33
  """Load a depth dataset by name. See :mod:`depth_estimation.data` for details."""
32
34
  from .data import load_dataset as _load
33
- return _load(name, split=split, root=root, download=download, transform=transform, **kwargs)
35
+
36
+ return _load(
37
+ name, split=split, root=root, download=download, transform=transform, **kwargs
38
+ )
34
39
 
35
40
 
36
41
  def __getattr__(name):
37
42
  """Defer torch-heavy imports until first use."""
38
43
  if name == "BaseDepthModel":
39
44
  from .modeling_utils import BaseDepthModel
45
+
40
46
  globals()["BaseDepthModel"] = BaseDepthModel
41
47
  return BaseDepthModel
42
48
  if name == "DepthProcessor":
43
49
  from .processing_utils import DepthProcessor
50
+
44
51
  globals()["DepthProcessor"] = DepthProcessor
45
52
  return DepthProcessor
46
53
  if name == "DepthPipeline":
47
54
  from .pipeline_utils import DepthPipeline
55
+
48
56
  globals()["DepthPipeline"] = DepthPipeline
49
57
  return DepthPipeline
50
58
  if name == "pipeline":
51
59
  from .pipeline_utils import pipeline
60
+
52
61
  globals()["pipeline"] = pipeline
53
62
  return pipeline
54
63
  # Training symbols
55
64
  if name == "DepthTrainer":
56
65
  from .trainer import DepthTrainer
66
+
57
67
  globals()["DepthTrainer"] = DepthTrainer
58
68
  return DepthTrainer
59
69
  if name == "DepthTrainingArguments":
60
70
  from .training_args import DepthTrainingArguments
71
+
61
72
  globals()["DepthTrainingArguments"] = DepthTrainingArguments
62
73
  return DepthTrainingArguments
63
74
  if name == "CombinedDepthLoss":
64
75
  from .losses import CombinedDepthLoss
76
+
65
77
  globals()["CombinedDepthLoss"] = CombinedDepthLoss
66
78
  return CombinedDepthLoss
67
79
  if name == "ScaleInvariantLoss":
68
80
  from .losses import ScaleInvariantLoss
81
+
69
82
  globals()["ScaleInvariantLoss"] = ScaleInvariantLoss
70
83
  return ScaleInvariantLoss
71
84
  if name == "GradientLoss":
72
85
  from .losses import GradientLoss
86
+
73
87
  globals()["GradientLoss"] = GradientLoss
74
88
  return GradientLoss
75
89
  if name == "BerHuLoss":
76
90
  from .losses import BerHuLoss
91
+
77
92
  globals()["BerHuLoss"] = BerHuLoss
78
93
  return BerHuLoss
79
94
  if name == "get_train_transforms":
80
95
  from .data.transforms import get_train_transforms
96
+
81
97
  globals()["get_train_transforms"] = get_train_transforms
82
98
  return get_train_transforms
83
99
  if name == "get_val_transforms":
84
100
  from .data.transforms import get_val_transforms
101
+
85
102
  globals()["get_val_transforms"] = get_val_transforms
86
103
  return get_val_transforms
104
+ # Video & streaming
105
+ if name == "VideoStream":
106
+ from .video import VideoStream
107
+
108
+ globals()["VideoStream"] = VideoStream
109
+ return VideoStream
110
+ if name == "process_video":
111
+ from .video import process_video
112
+
113
+ globals()["process_video"] = process_video
114
+ return process_video
115
+ # Visualization toolkit
116
+ if name == "viz":
117
+ from . import viz
118
+
119
+ globals()["viz"] = viz
120
+ return viz
121
+ # ONNX export
122
+ if name == "export_onnx":
123
+ from .export import export_onnx
124
+
125
+ globals()["export_onnx"] = export_onnx
126
+ return export_onnx
127
+ # Pruning
128
+ if name == "prune_model":
129
+ from .pruning import prune_model
130
+
131
+ globals()["prune_model"] = prune_model
132
+ return prune_model
133
+ if name == "compute_sparsity":
134
+ from .pruning import compute_sparsity
135
+
136
+ globals()["compute_sparsity"] = compute_sparsity
137
+ return compute_sparsity
138
+ if name == "make_pruning_permanent":
139
+ from .pruning import make_pruning_permanent
140
+
141
+ globals()["make_pruning_permanent"] = make_pruning_permanent
142
+ return make_pruning_permanent
143
+ # Quantization
144
+ if name == "quantize_model":
145
+ from .quantization import quantize_model
146
+
147
+ globals()["quantize_model"] = quantize_model
148
+ return quantize_model
149
+ if name == "quantize_onnx":
150
+ from .quantization import quantize_onnx
151
+
152
+ globals()["quantize_onnx"] = quantize_onnx
153
+ return quantize_onnx
87
154
  raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
88
155
 
89
156
 
@@ -107,6 +174,27 @@ __all__ = [
107
174
  "BerHuLoss",
108
175
  "get_train_transforms",
109
176
  "get_val_transforms",
177
+ # Video & streaming
178
+ "VideoStream",
179
+ "process_video",
180
+ # Visualization
181
+ "viz",
182
+ # ONNX export
183
+ "export_onnx",
184
+ # Pruning
185
+ "prune_model",
186
+ "compute_sparsity",
187
+ "make_pruning_permanent",
188
+ # Quantization
189
+ "quantize_model",
190
+ "quantize_onnx",
110
191
  ]
111
192
 
112
- __version__ = "0.0.9"
193
+ try:
194
+ from importlib.metadata import PackageNotFoundError, version
195
+
196
+ __version__ = version("depth_estimation")
197
+ except PackageNotFoundError:
198
+ # Package not installed (e.g. running from a source checkout without
199
+ # `pip install -e .`) — no metadata to read the version from.
200
+ __version__ = "0.0.0+unknown"