YOLOv5 训练边界框向左移动

Question

I'm trying to train my YOLOv5 to recognize car

That's a simple image in training folder

and inside my label i put this bounding box

i resize image and bounding box from 1000x750 to 640x640 and i start training my yolov5 with

.python train.py --img-size 640 --batch 16 --epochs 10 --data../dataset.yaml --weights yolov5s.pt --freeze 10

but the result from validation is this one

seems that all the coordinates of detected cars are shift to left

Any ideas?

Answer 1

i think you used bounding box in labeling tools.how about use labelme program. It is using outline instead of bounding box. if you detect car behind the tree, you should use outline instead of bounding box. beacause bounding box is more accuracy than outline. but,The detection is more contouring.