deep_sort_pytorch icon indicating copy to clipboard operation
deep_sort_pytorch copied to clipboard

Published 20 hours ago • verified publisher icon ZQPei

Wrongly tracked results

Open afesvas opened this issue 4 years ago • 3 comments

I changed a little of your code to track every object classes (80 classes in coco.names) and tested it with my own video.

In the results, some wrongly tracked objects as below are founded .

case 1. empty area is detected as an object results_0613 avi_010603133

case 2. one object is detected as two objects It happened when the detector confuses that the object is whether a car or truck. The object was classified as a truck at nth frame, but classified as a car at (n+1)th frame. Then at (n+1)th frame, the object was tracked as two objects. results_0613 avi_002328400

case 3. object is not detected There are 3 cars in the parking area, but only 2 cars are detected. 1

case 4. trackID is switched The left car was tracked with trackID 2329. but a few seconds later, changed to 2633. 2

Can anyone please let me know why these problems happen and how can I prevent them from happening? Or these problems are just accuracy limitations of detector and tracker models, so if I need more accurate results, should I use different models?

Thanks.

afesvas avatar Jun 19 '20 04:06 afesvas

Hi, I've also met similar problem on too many id switching despite I just track only 1 classes. I think that's the limitations of deepsort.

pvti avatar Jun 24 '20 15:06 pvti

You can do NMS to filter multiple observations associated with one target. And for the multiple ID Switches, you can use a pretrained Car re-identification appearance model.

derkbreeze avatar Aug 11 '20 10:08 derkbreeze

@afesvas @derkbreeze hello, i have question regarding reading the results.txt file, how to read it and what is the format of writting?

eng100200 avatar Sep 24 '20 08:09 eng100200