DerrickXuNu / OpenCOOD

[ICRA 2022] An opensource framework for cooperative detection. Official implementation for OPV2V.
https://mobility-lab.seas.ucla.edu/opv2v/
Other
663 stars 99 forks source link

Is the code of CIA-SSD complete? #63

Closed HuangZhe885 closed 1 year ago

HuangZhe885 commented 1 year ago

Thank you for your great work, I have a Q : Is the code of CIA-SSD complete? I'd want to discuss how D-NMS is implemented.

DerrickXuNu commented 1 year ago

Yes, I haven't gotten a chance to look into the problem you mentioned, as I am busy with my internship right now. Are you able to run FPV-RCNN code? Also, this part is implemented by @YuanYunshuang, so Yunshuang, maybe you are the best to help it.

YuanYunshuang commented 1 year ago

Thank you for your great work, I have a Q : Is the code of CIA-SSD complete? I'd want to discuss how D-NMS is implemented.

Hi, for CIASSD the scores for NMS is rectified with the predicted IOUs as suggested in the paper. You can find more details in the paper.

HuangZhe885 commented 1 year ago

Thank you for your great work, I have a Q : Is the code of CIA-SSD complete? I'd want to discuss how D-NMS is implemented.

Hi, for CIASSD the scores for NMS is rectified with the predicted IOUs as suggested in the paper. You can find more details in the paper.

I'm not sure whether it's my fault, but neither CIA-SSD nor FPV-RCNN will operate.

HuangZhe885 commented 1 year ago

Yes, I haven't gotten a chance to look into the problem you mentioned, as I am busy with my internship right now. Are you able to run FPV-RCNN code? Also, this part is implemented by @YuanYunshuang, so Yunshuang, maybe you are the best to help it.

I'm not sure whether it's my fault, but neither CIA-SSD nor FPV-RCNN will operate.

YuanYunshuang commented 1 year ago

Yes, I haven't gotten a chance to look into the problem you mentioned, as I am busy with my internship right now. Are you able to run FPV-RCNN code? Also, this part is implemented by @YuanYunshuang, so Yunshuang, maybe you are the best to help it.

I'm not sure whether it's my fault, but neither CIA-SSD nor FPV-RCNN will operate.

If you cannot compile it, it might be caused by the version problem. The newest version was upgraded to higher version. To spconv2 or over PyTorch 1.8. Some cuda code might not pass the compilation with PyTorch over 1.8.

HuangZhe885 commented 1 year ago

Yes, I haven't gotten a chance to look into the problem you mentioned, as I am busy with my internship right now. Are you able to run FPV-RCNN code? Also, this part is implemented by @YuanYunshuang, so Yunshuang, maybe you are the best to help it.

I'm not sure whether it's my fault, but neither CIA-SSD nor FPV-RCNN will operate.

If you cannot compile it, it might be caused by the version problem. The newest version was upgraded to higher version. To spconv2 or over PyTorch 1.8. Some cuda code might not pass the compilation with PyTorch over 1.8.

My spconv version is 2.0, and my pytorch version is 1.10.2+cu113, python =3.7. Additional algorithms can be run, however when these two are executed, an error is reported. when I run ciassd-early-fusion.yaml,it shows: File "/home/cood/OpenCOOD/opencood/models/sub_modules/mean_vfe.py", line 25, in forward voxel_features, voxel_num_points = batch_dict['voxel_features'], \ KeyError: 'voxel_features' when I run fpvrcnn_intermediate_fusion.yaml,it shows: File "/home/miniconda3/envs/COOD/lib/python3.7/site-packages/spconv/pytorch/conv.py", line 155, in _conv_forward 1] == self.in_channels, "channel size mismatch" AssertionError: channel size mismatch

YuanYunshuang commented 1 year ago

You might need to add the corresponding parameter for voxel feature and the correct input feature size in the yaml config file. Since I am recently traveling and not able to check the details of the errors but maybe you can try the fpvrcnn branch of this repo https://github.com/YuanYunshuang/OpenCOOD

This is the runnable old version. The recent updates of the official repo have many changes that might not support FPV-RCNN anymore and is not maintained anymore because this model is not very compatible with the other part of the code. And I recommend you to use spconv1.2.1 and PyTorch 1.8

HuangZhe885 commented 1 year ago

You might need to add the corresponding parameter for voxel feature and the correct input feature size in the yaml config file. Since I am recently traveling and not able to check the details of the errors but maybe you can try the fpvrcnn branch of this repo https://github.com/YuanYunshuang/OpenCOOD

This is the runnable old version. The recent updates of the official repo have many changes that might not support FPV-RCNN anymore and is not maintained anymore because this model is not very compatible with the other part of the code. And I recommend you to use spconv1.2.1 and PyTorch 1.8

Thank you for responding,I am extremely grateful! I'd want to pose the final question, which is contained in the paper: use IOU prediction to adjust confidence scores. I didn't discover the code for the exact implementation of this formula, and I didn't see the code for D-NMS. Is the requested API encapsulated?

YuanYunshuang commented 1 year ago

It is at line 91-93 of opencood/data_utils/post_processor/fpvrcnn_postprocessor.py

HuangZhe885 commented 1 year ago

It is at line 91-93 of opencood/data_utils/post_processor/fpvrcnn_postprocessor.py

Thank you for assisting me; you are really helpful to someone like me who is just starting started; thank you again!