Lecture 10 – Object Detection

HOG captures local edge/gradient structure by computing gradient orientations in small cells, accumulating them into histograms, and normalizing over blocks. This creates a descriptor robust to small deformations and illumination changes — ideal for pedestrian detection.

Pipeline: A fixed-size detection window (e.g., 64×128 pixels) slides across the image at multiple scales. At each position, HOG features are extracted and fed to a trained linear SVM. Positions scoring above the threshold are reported as detections. Overlapping detections are merged via NMS.

Selective Search starts with an over-segmentation and iteratively merges similar regions based on color, texture, size, and fill compatibility. The bounding boxes of all intermediate merged regions form the proposal set — typically 1000-2000 proposals covering objects at various scales.

R-CNN Pipeline: (1) Generate ~2000 region proposals via selective search. (2) Warp each proposal to a fixed size (e.g., 227×227). (3) Extract CNN features (e.g., AlexNet fc7). (4) Classify each proposal with class-specific SVMs. (5) Apply bounding box regression for refinement. This was the first method to combine deep features with region proposals for detection.

NMS removes redundant detections: (1) Sort all boxes by confidence. (2) Pick the top-scoring box, remove all boxes with IoU > threshold relative to it. (3) Repeat with remaining boxes until none are left. This ensures one detection per object.

IoU measures how much two bounding boxes overlap. It's the standard metric for evaluating detection accuracy (mAP uses IoU thresholds like 0.5 or 0.75) and for NMS suppression decisions.