ucas-vg / P2BNet

ECCV2022, Point-to-Box Network for Accurate Object Detection via Single Point Supervision

MIT License

63 stars 10 forks source link

strange result #3

Closed cxq1 closed 2 years ago

cxq1 commented 2 years ago

The following code was executed

[cmd 0] inference with trained P2BNet to get pseudo box

work_dir='../TOV_mmdetection_cache/work_dir/coco/' && CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=10000 tools/dist_train.sh configs2/COCO/P2BNet/P2BNet_r50_fpn_1x_coco_ms.py 8 \ --work-dir=${work_dir} \ --cfg-options evaluation.save_result_file=${work_dir}'_1200_latest_result.json' load_from=${work_dir}'P2BNet/epoch_12.pth' evaluation.do_first_eval=True runner.max_epochs=0

1661946453484 Wonder why the results are so bad

Darren-pfchen commented 2 years ago

Hello, the evaluation of P2BNet can be ignored. This is not the detection result. The val operation is to generate the result file： ${work_dir}'_1200_latest_result.json'，which transfer the point annotation to pseudo box annotation. Then they are used to train the detector like fastercnn, that is the final detection result. Sorry for missing the issue.

Darren-pfchen commented 2 years ago

Also, the ${work_dir}'_1200_latest_result.json file can be used for visualization, and the evaluation can be ignore (It means nothing, the gt_bboxes are hand-setting for mmdetection dataload).

cxq1 commented 2 years ago

Thank you for your reply