How set batch size in track.py? #930

kodinkod · 2023-06-07T09:50:46Z

kodinkod
Jun 7, 2023

how i may process video faster

Answered by mikel-brostrom

I guess this could be done. But the bottleneck is the reid model, not yolo. I will try to explain why

Yolo input: (bs, h,w,3)
Yolo output after NMS: [[N, 6], [M, 6], ..., [X, 6]]
Each of these output are used to crop each of the original input images. All the crops are then stacked together for each image and fed to the ReID model
ReID input : (N, h, w, 3), (M, h, w, 3), ...
The dimensions of a conv network are maximum 4. So you would have to process the images sequentially anyways for the ReID model. Not so simple as just increasing the batch size for yolo.

mikel-brostrom · 2023-06-07T19:06:41Z

I guess this could be done. But the bottleneck is the reid model, not yolo. I will try to explain why

Yolo input: (bs, h,w,3)
Yolo output after NMS: [[N, 6], [M, 6], ..., [X, 6]]
Each of these output are used to crop each of the original input images. All the crops are then stacked together for each image and fed to the ReID model
ReID input : (N, h, w, 3), (M, h, w, 3), ...
The dimensions of a conv network are maximum 4. So you would have to process the images sequentially anyways for the ReID model. Not so simple as just increasing the batch size for yolo.

0 replies