search

shaoshengsong / DeepSORT

support deepsort and bytetrack MOT(Multi-object tracking) using yolov5 with C++
765 stars 181 forks source link

Some code unfinished #11

Open taosean opened 3 years ago

taosean commented 3 years ago

Hi @shaoshengsong , thank you for your sharing of the C++ version of DeepSORT.

I checked the code and I found most of them no problem. However, I found something weird here, https://github.com/shaoshengsong/DeepSORT/blob/95a32f66c4d7271f6708e0bf71b52d7218606ca1/KalmanFilter/nn_matching.cpp#L133-L136 it seems the normalization of each row of the input 2 FEATURESS is not implemented.

In the repo of the original author of DeepSORT paper, the mentioned operation is implemented. https://github.com/nwojke/deep_sort/blob/280b8bdb255f223813ff4a8679f3e1321b08cdfc/deep_sort/nn_matching.py#L51-L53.

So, I want to know that this is not implemented on purpose or just by accident?

Thank you!

naserpiltan commented 3 years ago

Hi @taosean. I'm checking the code and I found something in the code that is not clear to me and I hope @shaoshengsong will give us an explaination about that. https://github.com/shaoshengsong/DeepSORT/blob/95a32f66c4d7271f6708e0bf71b52d7218606ca1/DeepAppearanceDescriptor/FeatureTensor.cpp#L120-L138

In line 131 , the author takes the image row pointers one by one : const uchar* inData = img.ptr<uchar>(i);

That's ok so far, but the image is not an grayscale image , its a RGB image and I think it should be like this : cv::Vec3b* inData = img.ptr<cv::Vec3b>(i); Whats your opinion? Thank you.