search

YutingXiao / Amodal-Segmentation-Based-on-Visible-Region-Segmentation-and-Shape-Prior

Implementation of the AAAI-2021 paper "Amodal Segmentation Based on Visible Region Segmentation and Shape Prior"
52 stars 7 forks source link

How to gain the performance of Mask RCNN on Visible region? #3

Closed zhixuanli closed 3 years ago

zhixuanli commented 3 years ago

  1. How to gain the performance of Mask RCNN on the Visible region?

  2. And after training the method of mask_rcnn_parallel_CtRef_VAR_SPRef_SPRet_FM_cocoacls_res50_SGD_1x, which terms should be considered as the final result on the performance table?

  3. If the answer is amodal ensemble and visible ensemble, then where's the performance of occluded area?

Thanks again for your patience!

YutingXiao commented 3 years ago

1.python tools/train_net.py --config-file configs/mask_rcnn_R_50_FPN_1x_visible.yaml.

2.amodal2_segm(refined) and visible2_segm(refined)

3.Forget about amodal_ensemble. it is nothing. The information of metrics can be found in "Installation"

YutingXiao commented 3 years ago

If you use cocoapi in the detectron2/data/amodal_datasets/pycocotools. The APs,APm,APl correspnonds to the performance on occluded objects. s: occlusion rate 0.15, m: occlusion rate 0.25, l: occlusion rate 0.4.

By the way, it hard to define the mAP(Occluded) since your visible prediction and amodal prediction may not be the same even if the object is not occluded. The metrics for evaluate performance are different on different papers. My design may not be perfect.

zhixuanli commented 3 years ago

Thanks a lot!!