search

sha2nkt / deco

Estimate vertex-level 3D human-scene and human-object contacts across the full body mesh
https://deco.is.tue.mpg.de/
Other
58 stars 4 forks source link

About using segmentation GT of 133 categories #11

Closed dqj5182 closed 8 months ago

dqj5182 commented 8 months ago

In the paper, it says "We obtain pseudo ground-truth scene segmentation masks, Xs ∈ R H×W×No , containing semantic labels for No = 133 categories". Do you modify the default option of Mask2Former to get 133 categories? If so, what modification do you impose to the base code?

dqj5182 commented 8 months ago

It seems that the only pretrained model with 133 categories is maskformer2_R50_bs16_50ep.yaml while the other issue, the DECO author mentions that they use maskformer2_swin_large_IN21k_384_bs16_100ep.yaml. May I ask which one was correct?

ac5113 commented 8 months ago

Please note that since coco_2017_val_panoptic has 133 class labels, if inference is performed using one of the panoptic segmentation models, the output would have 133 class labels. This is for all Mask2Former panoptic segmentation models on coco 2017 val panoptic.

You can check these 133 classes by following these steps:

  1. Install detectron2 python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
  2. from detectron2.data import MetadataCatalog coco_metadata = MetadataCatalog.get("coco_2017_val_panoptic") print(coco_metadata.stuff_classes)

Additionally, you can also obtain the corresponding colors using the stuff_colors key

dqj5182 commented 8 months ago

Thanks a lot!