Tracking without bells and whistles

 

https://arxiv.org/abs/1903.05625

 

The idea is very exciting and interesting.

It uses the previous Faster-RCNN without modification.

You have to understand Faster-RCNN first.

It has two stage detector. This Tacking model uses the second regression and classification models. The procedure is as below.

 

1. We needs a bounding box coordinated information of the previous box with the current frame.

2. Firstly, pass the bbox through the regression network and predict the next position.

3. Pass the nexy predicted bbox through the classification model. If the result indicates that the object is out of the image and occulated exceeding the threshold, the bbox's tracking should be killed.

4. Do detection with both regression and detection models. Then, compare each predicted boxes with previous tracked objects based on matching algorithms(euclidean or cosine distance). If there's no close object, a new tracker should be created.