search

VDIGPKU / CBNetV2

[TIP 2022] CBNetV2: A Composite Backbone Network Architecture for Object Detection
Apache License 2.0
369 stars 67 forks source link

Segmentation for HTC #4

Open heizie opened 2 years ago

heizie commented 2 years ago

Thanks for your work.

I was wondering, whether there are some bugs in your configs.

  1. you are using semantic_roi_extractor and semantic_head in your config for HTC. https://github.com/VDIGPKU/CBNetV2/blob/main/configs/cbnet/htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_adamw_20e_coco.py#L24
  2. and, you've used the original '../_base_/datasets/coco_instance.py',, which didn't provide the path of segmentation png. https://github.com/VDIGPKU/CBNetV2/blob/main/configs/_base_/datasets/coco_instance.py#L37

according to this issue: https://github.com/open-mmlab/mmdetection/issues/3767 with_seg should be True. But once you set this as True, then you should provide some segmentation png path for the training pipeline so how could you start the training?

tingtingliangvs commented 2 years ago

Thanks for your error report and we appreciate it a lot. The bug has been fixed in the latest version. Our HTC config follows htc_r50_fpn_1x_coco.py Due to limited computing resources, the training config of the large model has not been tested. If you have any other questions, please feel free to propose issues.