Bad results when running inference of annotation-free models

[argosdh]

Hi, thanks for your great work!

I'm encountering some problem when running the annotation-free setup models e.g. [RegionPLC + OpenScene + SparseUNet32] or [RegionPLC + SparseUNet32]. When inferenced on scannet validation, every point would be given the same label. This is my scripts used for running:
python3 test.py --cfg_file cfgs/scannet_models/spconv_clip_openvocab_test.yaml --ckpt ../checkpoints/sparseunet32_596.pth --save_results semantic

I barely changed the yaml file, only for those categories I want to add, like from
CATEGORY_NAMES: [door, window, desk, keyboard, others]
to
CATEGORY_NAMES: [trash can, television ,door, window, desk, keyboard, chair, others]

this is my yaml file, since I only need the qualitiative results, I put all my labels in base labels.

Thank you if you can reply !

[jihanyang]

Hello, you used the wrong config file here, you can follow the config file indicated in https://github.com/CVMI-Lab/PLA/blob/regionplc/docs/MODEL.md. Such as https://github.com/CVMI-Lab/PLA/blob/regionplc/tools/cfgs/scannet200_models/zs/spconv_clip_caption_openscene.yaml or https://github.com/CVMI-Lab/PLA/blob/regionplc/tools/cfgs/scannet200_models/zs/spconv_clip_caption.yaml

[argosdh]

Hello, you used the wrong config file here, you can follow the config file indicated in https://github.com/CVMI-Lab/PLA/blob/regionplc/docs/MODEL.md. Such as https://github.com/CVMI-Lab/PLA/blob/regionplc/tools/cfgs/scannet200_models/zs/spconv_clip_caption_openscene.yaml or https://github.com/CVMI-Lab/PLA/blob/regionplc/tools/cfgs/scannet200_models/zs/spconv_clip_caption.yaml

Thanks, I guess the my repo was the old one before the latest update. I'll try now.