search

zhenyuw16 / Uni3DETR

Code release for our NeurIPS 2023 paper "Uni3DETR: Unified 3D Detection Transformer", our ECCV 2024 paper "OV-Uni3DETR: Towards Unified Open-Vocabulary 3D Object Detection via Cycle-Modality Propagation"
Apache License 2.0
79 stars 3 forks source link

Missing clip_embed file and clarification on sunrgbd_coda data root #10

Closed rasd3 closed 2 weeks ago

rasd3 commented 2 weeks ago

I am trying to reproduce the results of OV-Uni3DETR but encountered an issue with the zeroshot_path(clip text embedding) file.

Specifically, the file referenced in zeroshot_path="clip_embed/sunrgbd_clip_a+cname_rn50_manyprompt_46c_coda.npy" is missing from the repository. I was unable to locate this .npy file in the provided code or any related documentation. Could you please provide guidance on where to find this file or how it can be generated?

Additionally, the data root for OV configuration is set to data/sunrgbd_coda. Could you clarify what "sunrgbd_coda" refers to and whether there are any specific instructions for setting it up? Any help on the proper dataset configuration or preprocessing steps would be greatly appreciated.

thank you

rasd3 commented 2 weeks ago

Is the file clip_embed/sunrgbd_clip_a+cname_rn50_manyprompt_46c_coda.npy generated from the notebook in this repository: https://github.com/yangcaoai/CoDA_NeurIPS2023/blob/main/CLIP/notebooks/Prompt_Engineering_for_ImageNet.ipynb?

rasd3 commented 2 weeks ago

I confirmed that the 'sunrgbd_clip_a+cname_rn50_manyprompt_46c_coda.npy' file can be created using the previous comment link, and I successfully reproduced the point-only performance during the experiment(9.11/48.4/18.25)