Seq-NMS_for_Video_Object_Detection.md

Seq-NMS for Video Object Detection

arXiv [TOC]

Introduction

1. difficulty:
   1. drastic scale changes
   2. occlusion
   3. motion blur

Contributions

1. Seq-NMS: the post-processing phase to use high-scoring object detections from nearby frames in order to boost scores of weaker detections within the same clip

Seq-NMS: 后处理过程，用临近帧的信息提升检测结果

Seq-NMS

Sequence Selection

find the maximum score sequence across the entire clip

Sequence Re-scoring

$S^{seq} = F(S^{seq'})$. We try two different re-scoring functions: the average and the max

Suppression

The boxes in the sequence are then removed from the set of boxes we link over. Furthermore, we apply suppression within frames such that if a bounding box in frame $t$; $t\in[t_s; t_e]$, has an IoU with bt over some threshold, it is also removed from the set of candidate boxes.