chengdianJiang
chengdianJiang
The train.py can work, but tensorboard shows that the loss is also zero, as shown in below....  Can you do me a favour? thanks
我也遇到了这种问题，我train和tracking时候用的图片统一成416 416，结果检测的效果很差，基本没有检测对的上的，训练48轮，刚开始loss收敛不错，40轮后loss小于0 了，作者可以给出一些解决思路吗？
The MOT label files can not be used, I made a script to convert txt to csv format, then it can work.
You can directly save the .txt file as .csv file, this is easy, but you need to modify two places according to this project. 1)change the img name to the...
Thanks for your reply, I trained the model with MOT17, in the RandomSampleCrop, The Croped imgs have been resized to 416*416, and the total loss was around~ 1.1, but the...