NMS filters out bounding boxes based on overlap and their confidence scores. The algorithm is as follows.
- Select proposal with highest confidence, remove from initial proposal list and add it to final proposal list.
- Compare this proposal with all others, remove those with IoU over threshold.
- Repeat above until there are no proposals left.
Soft NMS
SoftNMS uses a more relaxed policy for removing proposals. Instead of directly deleting proposals with high IoU, we can reduce their confidence score:
This increases precision significantly by preventing overlapping objects from interfering with each others’ predictions.