Open Zigars opened 3 years ago
my test is over, and these are my test results, I don't know which part have different, my VisDrone-yolov4 Rep is forked by yolov5's latest version, maybe the calculate of mAP have different?
Calculation of Precision and Recall are different, yolov5 calculate average score of 0.5:0.95 and yolor calculate score of 0.5. And I think AP(0.5) and AP(0.5:0.95) are using same calculation.
If that your explain is true, I think yolov4-p6-light.yaml (without reOrg and IDetect module) can not catch the yolov4-csp.yaml results in VisDrone dataset, although yolov4-p6-light.yaml have four output, the same infer time as yolov4-csp.yaml. I'm confuse about it. thank you for your explain! these are my yolov4-p6-light.yaml and yolov4-csp.yaml cfg files.yolov4-csp is a great work, I'm trying to make it more applicable for the VisDrone dataset. VisDrone-yolov4-cfg.zip
In my experiments. input resolution: performance 1280: p6 > p7 > p5 1536: p7 > p6 > p5
maybe for 640 case, p5 models will get better performance than p6 models, but i have not tested it yet.
thank you so much for your reply, I will test it in my code soon.
Zigars,我可以加你的QQ或者微信吗?我是一个学生,我想用自己的数据集跑作者的YOLOR,但是没跑通,可以请教下您吗?
I have test trained yolor-p6.yaml In VisDrone dataset (a famous UAV dataset) use 640 size and your pretrained .pt. And I get a excellent results, It's a great work! Then, I trained yolov4-p6-light.yaml(remove the reOrg and IDetect module and I fix it in yolov5's rep) in VisDrone dataset use 640 size without pretrained, but the results under the original yolov4-csp‘s result. maybe it have some bug in my code. So I'm training the same yolov4-p6-light.yaml(remove the reOrg and IDetect module and I fix it in your yolor's rep) in VisDrone use 640 size without pretrained, and I need find the bug in my own code if your yolor rep can get a good results. If not, that says yolov4-p6-light need a coco pretrained? because the model have four output, the loss can hard to convergence? The all experiment trained in 300epoch, 32batchsize, 640*640 size, use single V100 to train the model. maybe you can solve my question. thks!