pyfaceau 1.0.6__tar.gz → 1.0.9__tar.gz

{pyfaceau-1.0.6/pyfaceau.egg-info → pyfaceau-1.0.9}/PKG-INFO

@@ -1,15 +1,15 @@
 Metadata-Version: 2.4
 Name: pyfaceau
-Version: 1.0.6
+Version: 1.0.9
 Summary: Pure Python OpenFace 2.2 AU extraction with CLNF landmark refinement
 Home-page: https://github.com/johnwilsoniv/face-analysis
 Author: John Wilson
 Author-email: 
 License: CC BY-NC 4.0
-Project-URL: Homepage, https://github.com/johnwilsoniv/face-analysis
-Project-URL: Documentation, https://github.com/johnwilsoniv/face-analysis/tree/main/S0%20PyfaceAU
-Project-URL: Repository, https://github.com/johnwilsoniv/face-analysis
-Project-URL: Bug Tracker, https://github.com/johnwilsoniv/face-analysis/issues
+Project-URL: Homepage, https://github.com/johnwilsoniv/pyfaceau
+Project-URL: Documentation, https://github.com/johnwilsoniv/pyfaceau
+Project-URL: Repository, https://github.com/johnwilsoniv/pyfaceau
+Project-URL: Bug Tracker, https://github.com/johnwilsoniv/pyfaceau/issues
 Keywords: facial-action-units,openface,computer-vision,facial-analysis,emotion-recognition
 Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Science/Research
@@ -74,8 +74,14 @@ pyfaceau is a Python reimplementation of the [OpenFace 2.2](https://github.com/T
 #### Option 1: Install from PyPI (Recommended)
 
 ```bash
-# Install pyfaceau
-pip install pyfaceau
+# For NVIDIA GPU (CUDA):
+pip install pyfaceau[cuda]
+
+# For Apple Silicon (M1/M2/M3):
+pip install pyfaceau[coreml]
+
+# For CPU-only:
+pip install pyfaceau[cpu]
 
 # Download model weights (14MB)
 python -m pyfaceau.download_weights
@@ -84,12 +90,14 @@ python -m pyfaceau.download_weights
 # https://github.com/johnwilsoniv/face-analysis/tree/main/S0%20PyfaceAU/weights
 ```
 
+**Note:** PyFaceAU v1.1.0+ uses [PyMTCNN](https://pypi.org/project/pymtcnn/) for cross-platform face detection with CUDA/CoreML/CPU support.
+
 #### Option 2: Install from Source
 
 ```bash
 # Clone repository
 git clone https://github.com/johnwilsoniv/face-analysis.git
-cd "face-analysis/S0 PyfaceAU"
+cd "face-analysis/pyfaceau"
 
 # Install in development mode
 pip install -e .
@@ -104,13 +112,13 @@ pip install -e .
 ```python
 from pyfaceau import ParallelAUPipeline
 
-# Initialize parallel pipeline 
+# Initialize parallel pipeline
 pipeline = ParallelAUPipeline(
-    retinaface_model='weights/retinaface_mobilenet025_coreml.onnx',
     pfld_model='weights/pfld_cunjian.onnx',
     pdm_file='weights/In-the-wild_aligned_PDM_68.txt',
     au_models_dir='path/to/AU_predictors',
     triangulation_file='weights/tris_68_full.txt',
+    mtcnn_backend='auto',  # or 'cuda', 'coreml', 'cpu'
     num_workers=6,  # Adjust based on CPU cores
     batch_size=30
 )
@@ -125,20 +133,19 @@ print(f"Processed {len(results)} frames")
 # Typical output: ~28-50 FPS depending on CPU cores
 ```
 
-#### Standard Mode (4.6 FPS)
+#### Standard Mode (5-35 FPS depending on backend)
 
 ```python
 from pyfaceau import FullPythonAUPipeline
 
 # Initialize standard pipeline
 pipeline = FullPythonAUPipeline(
-    retinaface_model='weights/retinaface_mobilenet025_coreml.onnx',
     pfld_model='weights/pfld_cunjian.onnx',
     pdm_file='weights/In-the-wild_aligned_PDM_68.txt',
     au_models_dir='path/to/AU_predictors',
     triangulation_file='weights/tris_68_full.txt',
+    mtcnn_backend='auto',  # Automatically selects best backend
     use_calc_params=True,
-    use_coreml=True,  # macOS only
     verbose=False
 )
 
@@ -149,6 +156,11 @@ results = pipeline.process_video(
 )
 ```
 
+**Performance by backend:**
+- CUDA (NVIDIA GPU): ~35 FPS
+- CoreML (Apple Silicon): ~20-25 FPS
+- CPU: ~5-10 FPS
+
 ### Example Output
 
 ```csv
@@ -166,7 +178,7 @@ pyfaceau replicates the complete OpenFace 2.2 AU extraction pipeline:
 ```
 Video Input
     ↓
-Face Detection (RetinaFace ONNX)
+Face Detection (PyMTCNN - CUDA/CoreML/CPU)
     ↓
 Landmark Detection (PFLD 68-point)
     ↓

{pyfaceau-1.0.6 → pyfaceau-1.0.9}/README.md

@@ -29,8 +29,14 @@ pyfaceau is a Python reimplementation of the [OpenFace 2.2](https://github.com/T
 #### Option 1: Install from PyPI (Recommended)
 
 ```bash
-# Install pyfaceau
-pip install pyfaceau
+# For NVIDIA GPU (CUDA):
+pip install pyfaceau[cuda]
+
+# For Apple Silicon (M1/M2/M3):
+pip install pyfaceau[coreml]
+
+# For CPU-only:
+pip install pyfaceau[cpu]
 
 # Download model weights (14MB)
 python -m pyfaceau.download_weights
@@ -39,12 +45,14 @@ python -m pyfaceau.download_weights
 # https://github.com/johnwilsoniv/face-analysis/tree/main/S0%20PyfaceAU/weights
 ```
 
+**Note:** PyFaceAU v1.1.0+ uses [PyMTCNN](https://pypi.org/project/pymtcnn/) for cross-platform face detection with CUDA/CoreML/CPU support.
+
 #### Option 2: Install from Source
 
 ```bash
 # Clone repository
 git clone https://github.com/johnwilsoniv/face-analysis.git
-cd "face-analysis/S0 PyfaceAU"
+cd "face-analysis/pyfaceau"
 
 # Install in development mode
 pip install -e .
@@ -59,13 +67,13 @@ pip install -e .
 ```python
 from pyfaceau import ParallelAUPipeline
 
-# Initialize parallel pipeline 
+# Initialize parallel pipeline
 pipeline = ParallelAUPipeline(
-    retinaface_model='weights/retinaface_mobilenet025_coreml.onnx',
     pfld_model='weights/pfld_cunjian.onnx',
     pdm_file='weights/In-the-wild_aligned_PDM_68.txt',
     au_models_dir='path/to/AU_predictors',
     triangulation_file='weights/tris_68_full.txt',
+    mtcnn_backend='auto',  # or 'cuda', 'coreml', 'cpu'
     num_workers=6,  # Adjust based on CPU cores
     batch_size=30
 )
@@ -80,20 +88,19 @@ print(f"Processed {len(results)} frames")
 # Typical output: ~28-50 FPS depending on CPU cores
 ```
 
-#### Standard Mode (4.6 FPS)
+#### Standard Mode (5-35 FPS depending on backend)
 
 ```python
 from pyfaceau import FullPythonAUPipeline
 
 # Initialize standard pipeline
 pipeline = FullPythonAUPipeline(
-    retinaface_model='weights/retinaface_mobilenet025_coreml.onnx',
     pfld_model='weights/pfld_cunjian.onnx',
     pdm_file='weights/In-the-wild_aligned_PDM_68.txt',
     au_models_dir='path/to/AU_predictors',
     triangulation_file='weights/tris_68_full.txt',
+    mtcnn_backend='auto',  # Automatically selects best backend
     use_calc_params=True,
-    use_coreml=True,  # macOS only
     verbose=False
 )
 
@@ -104,6 +111,11 @@ results = pipeline.process_video(
 )
 ```
 
+**Performance by backend:**
+- CUDA (NVIDIA GPU): ~35 FPS
+- CoreML (Apple Silicon): ~20-25 FPS
+- CPU: ~5-10 FPS
+
 ### Example Output
 
 ```csv
@@ -121,7 +133,7 @@ pyfaceau replicates the complete OpenFace 2.2 AU extraction pipeline:
 ```
 Video Input
     ↓
-Face Detection (RetinaFace ONNX)
+Face Detection (PyMTCNN - CUDA/CoreML/CPU)
     ↓
 Landmark Detection (PFLD 68-point)
     ↓

pyfaceau-1.0.9/pyfaceau/clnf/__init__.py

@@ -0,0 +1,20 @@
+"""
+Python implementation of Constrained Local Neural Fields (CLNF).
+
+This package provides a pure Python implementation of the CLNF landmark
+detection algorithm for handling challenging cases like surgical markings
+and severe facial paralysis.
+
+Components:
+- cen_patch_experts: CEN patch expert loader and inference
+- pdm: Point Distribution Model for shape-constrained fitting
+- nu_rlms: Non-Uniform Regularized Landmark Mean Shift optimization
+- clnf_detector: Main CLNF detector integrating all components
+"""
+
+from .cen_patch_experts import CENPatchExperts
+from .pdm import PointDistributionModel
+from .nu_rlms import NURLMSOptimizer
+from .clnf_detector import CLNFDetector
+
+__all__ = ['CENPatchExperts', 'PointDistributionModel', 'NURLMSOptimizer', 'CLNFDetector']

pyfaceau-1.0.9/pyfaceau/clnf/cen_patch_experts.py

@@ -0,0 +1,439 @@
+#!/usr/bin/env python3
+"""
+CEN (Convolutional Expert Network) patch expert loader and inference.
+
+Loads and runs patch expert models from OpenFace 2.2's .dat format.
+"""
+
+import numpy as np
+import cv2
+from pathlib import Path
+import struct
+
+# Try to import numba for JIT compilation (optional)
+try:
+    from numba import njit
+    NUMBA_AVAILABLE = True
+except ImportError:
+    NUMBA_AVAILABLE = False
+    # Dummy decorator if numba not available
+    def njit(*args, **kwargs):
+        def decorator(func):
+            return func
+        return decorator if not args else decorator(args[0])
+
+
+class CENPatchExpert:
+    """
+    Single CEN patch expert for one landmark at one scale.
+
+    A patch expert is a small neural network that evaluates how likely
+    a landmark is at each position in a local patch.
+    """
+
+    def __init__(self):
+        self.width_support = 0
+        self.height_support = 0
+        self.weights = []  # List of weight matrices for each layer
+        self.biases = []   # List of bias vectors for each layer
+        self.activation_function = []  # Activation type for each layer
+        self.confidence = 0.0
+        self.is_empty = False
+
+    @classmethod
+    def from_stream(cls, stream):
+        """
+        Load a CEN patch expert from binary stream.
+
+        Args:
+            stream: Binary file stream positioned at patch expert data
+
+        Returns:
+            CENPatchExpert instance
+        """
+        expert = cls()
+
+        # Read header
+        read_type = struct.unpack('i', stream.read(4))[0]
+        if read_type != 6:
+            raise ValueError(f"Invalid CEN patch expert type: {read_type}, expected 6")
+
+        # Read dimensions and layer count
+        expert.width_support = struct.unpack('i', stream.read(4))[0]
+        expert.height_support = struct.unpack('i', stream.read(4))[0]
+        num_layers = struct.unpack('i', stream.read(4))[0]
+
+        # Empty patch (landmark invisible at this orientation)
+        if num_layers == 0:
+            expert.confidence = struct.unpack('d', stream.read(8))[0]
+            expert.is_empty = True
+            return expert
+
+        # Read layers
+        for i in range(num_layers):
+            # Activation function type
+            neuron_type = struct.unpack('i', stream.read(4))[0]
+            expert.activation_function.append(neuron_type)
+
+            # Read bias matrix
+            bias = read_mat_bin(stream)
+            expert.biases.append(bias)
+
+            # Read weight matrix
+            weight = read_mat_bin(stream)
+            expert.weights.append(weight)
+
+        # Read confidence
+        expert.confidence = struct.unpack('d', stream.read(8))[0]
+
+        return expert
+
+    def response(self, area_of_interest):
+        """
+        Compute patch expert response map for an image patch.
+
+        Args:
+            area_of_interest: Grayscale image patch (H, W) as float32
+
+        Returns:
+            response: Response map (response_height, response_width) as float32
+        """
+        if self.is_empty:
+            # Return zero response for empty patches
+            response_height = max(1, area_of_interest.shape[0] - self.height_support + 1)
+            response_width = max(1, area_of_interest.shape[1] - self.width_support + 1)
+            return np.zeros((response_height, response_width), dtype=np.float32)
+
+        # Apply contrast normalization (row-wise)
+        normalized = contrast_norm(area_of_interest)
+
+        # Convert to column format with bias (im2col)
+        input_col = im2col_bias(normalized, self.width_support, self.height_support)
+
+        # Forward pass through neural network layers
+        layer_output = input_col
+        for i in range(len(self.weights)):
+            # Linear: output = input * weight^T + bias
+            layer_output = layer_output @ self.weights[i].T + self.biases[i]
+
+            # Apply activation function
+            if self.activation_function[i] == 0:
+                # Sigmoid (clamp extreme values to prevent overflow)
+                layer_output = np.clip(layer_output, -88, 88)
+                layer_output = 1.0 / (1.0 + np.exp(-layer_output))
+            elif self.activation_function[i] == 1:
+                # Tanh
+                layer_output = np.tanh(layer_output)
+            elif self.activation_function[i] == 2:
+                # ReLU
+                layer_output = np.maximum(0, layer_output)
+            # else: linear (no activation)
+
+        # Reshape output to 2D response map
+        response_height = area_of_interest.shape[0] - self.height_support + 1
+        response_width = area_of_interest.shape[1] - self.width_support + 1
+
+        # Output is (num_patches,) -> reshape to (height, width)
+        # Patches are now in row-major order (natural numpy order)
+        response = layer_output.reshape(response_height, response_width)
+
+        return response.astype(np.float32)
+
+
+class CENPatchExperts:
+    """
+    Collection of CEN patch experts for all landmarks and scales.
+
+    Manages multi-scale patch experts from OpenFace 2.2 .dat files.
+    """
+
+    def __init__(self, model_dir):
+        """
+        Load CEN patch experts from model directory.
+
+        Args:
+            model_dir: Path to directory containing cen_patches_*.dat files
+        """
+        self.model_dir = Path(model_dir)
+        self.patch_scaling = [0.25, 0.35, 0.50, 1.00]  # Scales available
+        self.num_landmarks = 68
+
+        # patch_experts[scale][landmark]
+        self.patch_experts = []
+
+        # Load all scale levels
+        print(f"Loading CEN patch experts (410 MB, ~5-10 seconds)...")
+        for idx, scale in enumerate(self.patch_scaling):
+            scale_file = self.model_dir / f"cen_patches_{scale:.2f}_of.dat"
+            if not scale_file.exists():
+                # Try patch_experts subdirectory
+                scale_file = self.model_dir / "patch_experts" / f"cen_patches_{scale:.2f}_of.dat"
+            if not scale_file.exists():
+                raise FileNotFoundError(f"CEN model not found: {scale_file}")
+
+            print(f"  [{idx+1}/4] Loading scale {scale}...")
+            experts_at_scale = self._load_scale(scale_file)
+            self.patch_experts.append(experts_at_scale)
+            print(f"      ✓ {len(experts_at_scale)} patch experts loaded")
+
+    def _load_scale(self, dat_file):
+        """
+        Load all patch experts at one scale level.
+
+        Args:
+            dat_file: Path to .dat file
+
+        Returns:
+            List of CENPatchExpert instances (frontal view only)
+        """
+        with open(dat_file, 'rb') as f:
+            # File structure (confirmed from OpenFace C++ source):
+            # 1. Header: patch_scale (double) + num_views (int)
+            # 2. View centers: num_views × (x, y, z) as doubles
+            # 3. Visibility matrices: num_views × cv::Mat<int>
+            # 4. Patch experts: num_views × num_landmarks × CEN_patch_expert
+
+            # Read file header
+            patch_scale = struct.unpack('d', f.read(8))[0]
+            num_views = struct.unpack('i', f.read(4))[0]
+
+            # Read view centers (3D orientation for each view)
+            # Each view center is immediately followed by an empty matrix
+            view_centers = []
+            for _ in range(num_views):
+                x = struct.unpack('d', f.read(8))[0]
+                y = struct.unpack('d', f.read(8))[0]
+                z = struct.unpack('d', f.read(8))[0]
+                view_centers.append((x, y, z))
+
+                # Read empty matrix immediately after this view center
+                empty_mat = read_mat_bin(f)  # Should be 0×0 matrix
+
+            # Read visibility matrices (one per view)
+            # These are cv::Mat<int> storing which landmarks are visible in each view
+            visibilities = []
+            for _ in range(num_views):
+                vis_mat = read_mat_bin(f)
+                visibilities.append(vis_mat)
+
+            # Read mirror metadata (for facial symmetry)
+            mirror_inds = read_mat_bin(f)  # 1×68 matrix
+            mirror_views = read_mat_bin(f)  # 1×7 matrix
+
+            # Now read patch experts for all views
+            all_experts = []
+            for view_idx in range(num_views):
+                experts_for_view = []
+                for lm_idx in range(self.num_landmarks):
+                    try:
+                        expert = CENPatchExpert.from_stream(f)
+                        experts_for_view.append(expert)
+                    except Exception as e:
+                        print(f"Error loading expert (view={view_idx}, landmark={lm_idx}): {e}")
+                        raise
+                all_experts.append(experts_for_view)
+
+        # Return only frontal view (view 0) for now
+        # TODO: Add multi-view support for profile faces
+        return all_experts[0]
+
+    def response(self, image, landmarks, scale_idx):
+        """
+        Compute patch expert responses for all landmarks at given scale.
+
+        Args:
+            image: Grayscale image (H, W) as float32 [0, 255]
+            landmarks: Current landmark positions (68, 2) as float32
+            scale_idx: Scale index (0-3)
+
+        Returns:
+            responses: List of 68 response maps (one per landmark)
+            extraction_bounds: List of 68 tuples (x1, y1, x2, y2) with actual extraction bounds
+        """
+        if scale_idx < 0 or scale_idx >= len(self.patch_experts):
+            raise ValueError(f"Invalid scale index: {scale_idx}")
+
+        experts_at_scale = self.patch_experts[scale_idx]
+        responses = []
+        extraction_bounds = []
+
+        # For each landmark, extract patch and compute response
+        for lm_idx in range(self.num_landmarks):
+            expert = experts_at_scale[lm_idx]
+
+            # Extract a SEARCH AREA around the landmark
+            # The search area should be larger than the support window to allow
+            # the patch expert to evaluate multiple positions
+            # Search radius: 2.0x support (OpenFace default for robust refinement)
+            search_radius = int(max(expert.width_support, expert.height_support) * 2.0)
+
+            lm_x, lm_y = landmarks[lm_idx]
+            x1 = max(0, int(lm_x - search_radius))
+            y1 = max(0, int(lm_y - search_radius))
+            x2 = min(image.shape[1], int(lm_x + search_radius))
+            y2 = min(image.shape[0], int(lm_y + search_radius))
+
+            # Extract and compute response
+            patch = image[y1:y2, x1:x2]
+            if patch.size > 0 and patch.shape[0] > expert.height_support and patch.shape[1] > expert.width_support:
+                response = expert.response(patch)
+            else:
+                response = np.zeros((1, 1), dtype=np.float32)
+
+            responses.append(response)
+            extraction_bounds.append((x1, y1, x2, y2))
+
+        return responses, extraction_bounds
+
+
+def read_mat_bin(stream):
+    """
+    Read OpenCV matrix from binary stream (OpenFace format).
+
+    Args:
+        stream: Binary file stream
+
+    Returns:
+        numpy array with matrix data
+    """
+    # Read dimensions and type
+    rows = struct.unpack('i', stream.read(4))[0]
+    cols = struct.unpack('i', stream.read(4))[0]
+    cv_type = struct.unpack('i', stream.read(4))[0]
+
+    # Handle empty matrices (0×0 or 0×N or N×0)
+    if rows == 0 or cols == 0:
+        return np.array([], dtype=np.float32).reshape(rows, cols) if rows >= 0 and cols >= 0 else np.array([])
+
+    # Map OpenCV type to numpy dtype
+    # OpenCV type codes: CV_8U=0, CV_8S=1, CV_16U=2, CV_16S=3,
+    #                    CV_32S=4, CV_32F=5, CV_64F=6
+    if cv_type == 0:  # CV_8U
+        dtype = np.uint8
+    elif cv_type == 1:  # CV_8S
+        dtype = np.int8
+    elif cv_type == 2:  # CV_16U
+        dtype = np.uint16
+    elif cv_type == 3:  # CV_16S
+        dtype = np.int16
+    elif cv_type == 4:  # CV_32S
+        dtype = np.int32
+    elif cv_type == 5:  # CV_32F
+        dtype = np.float32
+    elif cv_type == 6:  # CV_64F
+        dtype = np.float64
+    else:
+        raise ValueError(f"Unsupported OpenCV matrix type: {cv_type} (rows={rows}, cols={cols})")
+
+    # Read data
+    size = rows * cols
+    data = np.frombuffer(stream.read(size * np.dtype(dtype).itemsize), dtype=dtype)
+
+    # Reshape to matrix (OpenCV uses row-major order like NumPy)
+    matrix = data.reshape(rows, cols)
+
+    # For weight/bias matrices, convert to float32; for visibility, keep as-is
+    if cv_type in [5, 6]:  # Float types
+        return matrix.astype(np.float32)
+    else:  # Integer types
+        return matrix
+
+
+@njit(fastmath=True, cache=True)
+def _contrast_norm_numba(input_patch, output):
+    """Numba-optimized contrast normalization (5-10x faster)."""
+    for y in range(input_patch.shape[0]):
+        # Skip first column, compute mean of rest
+        row_sum = 0.0
+        cols = input_patch.shape[1] - 1
+        for x in range(1, input_patch.shape[1]):
+            row_sum += input_patch[y, x]
+        mean = row_sum / cols
+
+        # Compute standard deviation
+        sum_sq = 0.0
+        for x in range(1, input_patch.shape[1]):
+            diff = input_patch[y, x] - mean
+            sum_sq += diff * diff
+
+        norm = np.sqrt(sum_sq)
+        if norm < 1e-10:
+            norm = 1.0
+
+        # Normalize (skip first column)
+        output[y, 0] = input_patch[y, 0]  # Keep first column
+        for x in range(1, input_patch.shape[1]):
+            output[y, x] = (input_patch[y, x] - mean) / norm
+
+    return output
+
+
+def contrast_norm(input_patch):
+    """
+    Apply row-wise contrast normalization.
+
+    Uses Numba JIT if available for 5-10x speedup.
+
+    Args:
+        input_patch: Image patch (H, W) as float32
+
+    Returns:
+        normalized: Contrast-normalized patch
+    """
+    output = np.empty_like(input_patch, dtype=np.float32)
+
+    if NUMBA_AVAILABLE:
+        return _contrast_norm_numba(input_patch, output)
+    else:
+        # Fallback: NumPy vectorized version
+        output = input_patch.copy()
+        for y in range(input_patch.shape[0]):
+            row = input_patch[y, 1:]  # Skip first column
+            mean = np.mean(row)
+            norm = np.std(row)
+            if norm < 1e-10:
+                norm = 1.0
+            output[y, 1:] = (input_patch[y, 1:] - mean) / norm
+        return output
+
+
+def im2col_bias(input_patch, width, height):
+    """
+    Convert image to column format with bias for convolutional processing.
+
+    Uses vectorized stride tricks for 10-20x speedup over loop-based version.
+
+    Args:
+        input_patch: Image patch (m, n) as float32
+        width: Sliding window width
+        height: Sliding window height
+
+    Returns:
+        output: Matrix (num_windows, width*height+1) with bias column
+    """
+    m, n = input_patch.shape
+    y_blocks = m - height + 1
+    x_blocks = n - width + 1
+    num_windows = y_blocks * x_blocks
+
+    # Use stride tricks to create view of all sliding windows (zero-copy!)
+    from numpy.lib.stride_tricks import as_strided
+
+    # Create 4D array: (y_blocks, x_blocks, height, width)
+    windows = as_strided(
+        input_patch,
+        shape=(y_blocks, x_blocks, height, width),
+        strides=(input_patch.strides[0], input_patch.strides[1],
+                input_patch.strides[0], input_patch.strides[1])
+    )
+
+    # Reshape windows - keep natural row-major order
+    # Original: windows shape is (y_blocks, x_blocks, height, width)
+    # Flatten to (num_windows, width*height) in row-major order (natural numpy order)
+    windows_flat = windows.reshape(num_windows, height * width)
+
+    # Add bias column
+    output = np.ones((num_windows, height * width + 1), dtype=np.float32)
+    output[:, 1:] = windows_flat
+
+    return output