2016 ICLR workshow

[1602.08465] Seq-NMS for Video Object Detection [PDF] [notes]

Wei Han, Pooya Khorrami, Tom Le Paine, Prajit Ramachandran, Mohammad Babaeizadeh, Honghui Shi, Jianan Li, Yan Shuicheng, ,Thomas S. Huang

read 07/03/2018

Objective

Simple method to perform object tracking in a video Strong detections in previous frames should boost weaker ones from other frames

Synthesis

Sequence selection

Frame-based image detector gives proposals with various scores for a given class Construct a graph where detections are linked if their IOU is > 0.5 In the graph, keep paths the maximize the sequence score (sum of individual detection score accross path of linked frames) accross the entire clip

Re-scoring

A new score is produced for the sequence by taking either the average or the max of the detection scores along the generated path ==> average performs better then max on ImageNet VID dataset

Suppression

The boxes in the best sequence are removed from the video Boxes in frames that have high overlap with the best-sequence detections are also removed