Learning-Deep-Learning

Adaptive NMS: Refining Pedestrian Detection in a Crowd

October 2020

tl;dr: Predict the object density (crowdedness) score and use it as adaptive threshold for NMS.

Overall impression

The paper has a simple intuition: that high NMS threshold keeps more crowded instances while a low NMS threshold wipes out more false positives. The key question is how to predict the crowdedness in inference time? The paper proposes to predict it together with other attributes of bboxes.

Both RepLoss and AggLoss proposes additional penaltiies to produce more compact bounding boxes and become less sensitive to NMS.

Key ideas

Technical details

Notes