environment problem

[FlyingPz]

Due to some reasons, Docker cannot be used and there have been many problems downloading through PIP. Could you please provide me with the environment txt to take a look? Thank you very much

[ziqipang]

@FlyingPz Thanks for the question! Unfortunately, I don't have access to the server since I have finished my internship. Does the second section in this document works for you? https://github.com/TRI-ML/PF-Track/blob/main/documents/environment.md

[FlyingPz]

Thank you for your answer. I found that I encountered issues with pip install mmdet3d. Following PETR's installation guide, I later used the git clone method to successfully compile the mmdet3d environment and got the code running successfully.

I have another question regarding the DETR3D and PETR detectors at a resolution of 1600*800: approximately how much GPU memory do they each consume? I am currently using a server with a 2080 Ti (12GB) and will later switch to a 4090 (24GB). Thank you very much for your help.

[ziqipang]

@FlyingPz I was using 40GB A100 at that time and 1600x800 was challenging. So I think 1600*800 would be challenging for you. But I encourage you to try or freeze some components.

[Johnny-Hu-406]

@FlyingPz I am a beginner and would like to ask if you were able to successfully complete training using an RTX 4090 (24GB) for this project? I’m currently working on a similar setup and wanted to know if the 4090 would be sufficient for smooth training.

Thank you very much for your help!

[FlyingPz]

@FlyingPz I am a beginner and would like to ask if you were able to successfully complete training using an RTX 4090 (24GB) for this project? I’m currently working on a similar setup and wanted to know if the 4090 would be sufficient for smooth training.

Thank you very much for your help!

It is feasible to train a single-frame detector first and then freeze the backbone network to train the tracker. I replaced the backbone network with ResNet101, which reduces the number of parameters and enables training on 4090