SSD object detection - notitiam/ML-paper-notes GitHub Wiki
- Image classification vs Object detection vs Segmentation
- Two stage approaches:
- Region proposal
- e.g. Sliding window approach
- Detection
- R-CNN, Fast R-CNN, Faster R-CNN
- Problems: Computationally expensive
- Single shot detectors: SSD, YOLO, YOLO v3 etc
Details:
- Multi scale feature map
- Default boxes and aspect ratios (Anchor boxes)
- Training:
- choose feature map
- location in feature map
- associated ground truth box
- train all boxes with IoU > 0.5
- hard negative mining: highest confidence loss (3:1, negative:positive)
- 11 priors per feature map cell, total of 1420 priors per image