IDEA-Research
Reproduction results
Dear authors, I trained the configs/coco/instance-segmentation/maskdino_R50_bs16_50ep_3s.yaml configuration directly with 8xa100 GPUs and obtained the following results, which differ significantly from your reported performance. Is this deviation reasonable?
[07/24 22:46:56] d2.engine.defaults INFO: Evaluation results for coco_2017_val in csv format:
[07/24 22:46:56] d2.evaluation.testing INFO: copypaste: Task: bbox
[07/24 22:46:56] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl
[07/24 22:46:56] d2.evaluation.testing INFO: copypaste: 50.9154,70.0297,55.9029,33.9335,54.3486,65.7561
[07/24 22:46:56] d2.evaluation.testing INFO: copypaste: Task: segm
[07/24 22:46:56] d2.evaluation.testing INFO: copypaste: AP,AP50,AP75,APs,APm,APl
[07/24 22:46:56] d2.evaluation.testing INFO: copypaste: 45.6611,68.4114,49.7517,25.5672,49.0397,65.0208
the repo is bbox map 51.5 mask map 46.1 but I got bbox map 50.9 mask map 45.7.
Hi, thanks for reaching out. The deviation is a bit large. Did you use the mask-guided anchor box initialization? Please clone our latest code and see if the setting is wrong.
As we did not run the model multiple times, we are not sure whether the deviation of about 0.5 is reasonable. If there is nothing wrong with your setting, you may run it again to see the results. You can also run this config to see the results.
@FengLi-ust Thank you for your reply. I did not modify your configuration, I trained directly with the default initialization method of bitmask.