PaperReading

Visibility Guided NMS: Efficient Boosting of Amodal Object Detection in Crowded Traffic Scenes

June 2020

tl;dr: Use modal bbox to guide NMS of amodal bbox.

Overall impression

The paper addressed a critical issue for autonomous driving in parking lot and in urban areas where many parked cars are heavily occluded. Same issue happens to crowds of pedestrians as well.

Humans perceive the world by predicting the bbox covering the entire object even if it is occluded. This is called amodal perception. (cf. Amodal completion)

This is very similar to R2 NMS in CVPR 2020, which focuses on crowd pedestrian detection.

Key ideas

Technical details

Notes