Closed tenggyut closed 2 years ago
@tenggyut we don't provide support for custom code.
Repo mAP is typically lower than pycocotools mAP, i.e. see https://github.com/ultralytics/yolov5/issues/2258
YOLOv5s VOC is about 0.86mAP@0.05 https://wandb.ai/glenn-jocher/VOC
problem found! this repo ignores difficult ground truthes which I haven't. I'll close this issue.
Thanks, I get it.
@tenggyut we don't provide support for custom code.
Repo mAP is typically lower than pycocotools mAP, i.e. see #2258
YOLOv5s VOC is about 0.86mAP@0.05 https://wandb.ai/glenn-jocher/VOC
do u use VOC 2007 or VOC 2012 or both?
@NDTuong we use VOC 2007, VOC 2012, and VOC 2007+2012 combined for training. More training details can be found in the Ultralytics Docs: https://docs.ultralytics.com/yolov5/.
Search before asking
Question
I have used this repo(v6.0) to train yolov5s on voc dataset( 07+12 trainval as the training set, and 07 test as the val set), after 30+ epochs, the best mAP@0.5 is about 0.828 calculated by the eval function in val.py。
I then use the best checkpoint to run val.py and save the predictions in txt file, then I wrote a script to load the groundtruthes and the predictions and evaluated the predictions using pycocotools, the mAP@0.5 is about 0.77 which is far less than 0.828....
the commd I use to save predictions is
python val.py --data data/voc_rock.yaml --weights runs/train/exp16/weights/best.pt --conf-thres 0.01 --iou-thres 0.45 --save-txt --save-json --half --save-conf
the script I wrote to calcuate coco metric is
also I have found a comment that says yolov5s can achive 0.82 on voc2007 test set.
So am I missing something here?
Additional
No response