To perform Amodal Instance Segmentation

[masa-suke]

Thank you for your great work!

I am trying to train a model to perform Amodal Instance Segmentation and input my own images to see the prediction results of hidden regions. I would like to receive the following and other predicted results. However, I am unable to detect hidden regions. So, please let me ask some questions.

Questions

• About the label data

  • datasets/coco/annotations/instances_train_2017_transform_slight_correct.json Which of the following ABCs is appropriate? I did C but did not get good results.
    • A. Use a file downloaded from google drive or onedrive.
    • B. Run BCNet/process.h to the json file downloaded from google drive or onedrive and use the output file.
    • C. Run BCNet/process.h to the instance_train.json contained in the 2017 Train/Val annotations [241MB] and use the output file.
  • datasets/coco/annotations/instances_val2017.json I used the instances_val2017.json included in the 2017Train/Val annotation [241MB], Do I also need to run BCNet/process.h against instances_val2017.json?

• About config file I am currently running BCNet/all.sh and using the BCNet/configs/fcos/fcos_imprv_R_50_FPN.yaml listed in it as the training config. Would BCNet/configs/fcos/fcos_imprv_R_101_FPN.yaml be more appropriate?

• About visualize.sh Is the same setting as for `all.sh' correct? I am concerned that the config used for visualize.sh and all.sh is different.