# Object Detection ![[object detection.jpg]] ## Window-based Detection Sliding window detection and global appearance descriptors: - Simple detection protocol to implement - Good feature choices critical - Past successes for certain classes Problems - High computational complexity - For example: 250,000 locations x 30 orientations x 4 scales = 30,000,000 evaluations! - Many boxes - Many scales - Many box ratio's - If training binary detectors independently, means cost increases linearly with number of classes - With so many windows, false positive rate better be low ## Selective Search Based on the idea that images are hierarchical. Proposed by University of Amsterdam in IJCAI 2013. ![[selective-search.jpg]] Group segmenets on similarity: - Color - Texture ([[Scale-Invariant Feature Transform (SIFT)]] like) - Size (small regions merge earlier) - Fill (regions with small borders merge laters) Example: ![[selective-example.jpg]] ## Intersection over union There will be several similar bounding boxes, we use IoU to find the best box. The higher the better. ![[intersection-over-union.jpg]] ## Non-maximum Suppression In order to further reduce the number of boxes, sort the boxes by classification confidence, and remove the ones with overlap on these high confidence boxes. This technique is called non-maximum suppresion. --- ## References