search

ifzhang / ByteTrack

[ECCV 2022] ByteTrack: Multi-Object Tracking by Associating Every Detection Box
MIT License
4.69k stars 891 forks source link

跟踪的框偏大 #223

Open zzh805780186 opened 2 years ago

zzh805780186 commented 2 years ago

捕获 yolox-s检测后的框大小正常，bytetrack跟踪后的框就变得很大，这是什么情况呢？

zzh805780186 commented 2 years ago

这是对应的yolox-s的检测结果，检测的框大小比较正常 74

NanH5837 commented 2 years ago

Hello, my tracking output is much larger than you. The aspect ratio is greater than 1.6 so that it is filtered out. Have you found the cause of the problem? @zzh805780186

dyhBUPT commented 2 years ago

I think It is caused by the Kalman Filter.

zzh805780186 commented 2 years ago

是的，前一帧的时候因为目标只露出一半，所以检测框是扁的，下一帧卡尔曼预测错误，会更新出错误尺寸及比例的框

YuNaruto commented 2 years ago

@zzh805780186 @dyhBUPT 是的，我也有类似的情况，当一个物体从镜头边缘出现时，跟踪框都会先变大，然后变成正常。有什么好的解决方法吗？

dyhBUPT commented 2 years ago

@YuNaruto I have solved the problem by proposing the NSA Kalman Filter method. You can refer to the paper of GIAOTracker for more details and refer to the codes of StrongSORT for implementation.

Best wishes.

YuNaruto commented 2 years ago

@dyhBUPT

Thank you. I will read this paper carefully. Thank you again!

SEUvictor commented 2 years ago

@YuNaruto I have solved the problem by proposing the NSA Kalman Filter method. You can refer to the paper of GIAOTracker for more details and refer to the codes of StrongSORT for implementation.

Best wishes.

竟然是StrongSort的作者！大佬好！膜拜一下！ 想在此请教一下，想做一个基于yolov7的车辆跟踪，希望IDs不要太严重，然后尽量可以实时，您觉得yolov7配上bytetrack好呢，还是StrongSort会还一点呢？看了您的StrongSort，貌似fps比较低？ 感谢！

dyhBUPT commented 2 years ago

@YuNaruto I have solved the problem by proposing the NSA Kalman Filter method. You can refer to the paper of GIAOTracker for more details and refer to the codes of StrongSORT for implementation. Best wishes.

竟然是StrongSort的作者！大佬好！膜拜一下！ 想在此请教一下，想做一个基于yolov7的车辆跟踪，希望IDs不要太严重，然后尽量可以实时，您觉得yolov7配上bytetrack好呢，还是StrongSort会还一点呢？看了您的StrongSort，貌似fps比较低？ 感谢！

Thanks. You can try NSA Kalman Filter for your "big bbox" problem. It seems work well for @YuNaruto (https://github.com/dyhBUPT/StrongSORT/issues/41)

StrongSORT has a low FPS because of using the ReID model. So, ByteTrack or StrongSORT, it depends on your data.

Best wishes.

SEUvictor commented 2 years ago

@YuNaruto I have solved the problem by proposing the NSA Kalman Filter method. You can refer to the paper of GIAOTracker for more details and refer to the codes of StrongSORT for implementation. Best wishes.

竟然是StrongSort的作者！大佬好！膜拜一下！ 想在此请教一下，想做一个基于yolov7的车辆跟踪，希望IDs不要太严重，然后尽量可以实时，您觉得yolov7配上bytetrack好呢，还是StrongSort会还一点呢？看了您的StrongSort，貌似fps比较低？ 感谢！

Thanks. You can try NSA Kalman Filter for your "big bbox" problem. It seems work well for @YuNaruto (dyhBUPT/StrongSORT#41)

StrongSORT has a low FPS because of using the ReID model. So, ByteTrack or StrongSORT, it depends on your data.

Best wishes.

Thank you very much!