search

yangyangyang127 / PointCLIP_V2

[ICCV 2023] PointCLIP V2: Prompting CLIP and GPT for Powerful 3D Open-world Learning
MIT License
232 stars 14 forks source link

Zero-shot classification of ScanObjectNN #8

Open kaimingkuang opened 1 year ago

kaimingkuang commented 1 year ago

Thanks for your amazing work! I was trying to run zero-shot classification of ScanObjectNN using the pipeline you provided. However, I only got 29.78% accuracy for the OBJ_ONLY variation (50.09% reported in the paper) using the following script:

python main.py \
--trainer PointCLIPV2_ZS \
--dataset-config-file configs/datasets/scanobjectnn.yaml \
--config-file configs/trainers/PointCLIPV2_ZS/rn50.yaml \
--output-dir output/PointCLIPV2_ZS/rn50/scanobjectnn \
--no-train \
--zero-shot \
--test-file=data/scanobjectnn/main_split_nobg/test_objectdataset.h5  > obj_only_mainnobg.log

The test-file argument specifies which variation to use. Am I missing something here? Thank you!

yangyangyang127 commented 1 year ago

Sorry for the late reply.

Actually, the main backbone we use is the vit_b16 visual encoder, which achieves a higher accuracy than rn50 backbone. So you can try change the config file to vit_b16.yaml.

And thanks for your attention.