Low mAP when training Transfusion with lidar_only on nuscenes v1.0-mini dataset

[2460707925]

I am trying to reproduce the results of the Transfusion model on the nuscenes v1.0-mini dataset using the lidar_only configuration (transfusion_nusc_pillar_L.py). However, after training for 20 epochs on a single 3090ti GPU, my mAP is only 0.1611, which is much lower than the reported mAP in the paper.

I have followed the instructions in the official Transfusion repository and used the latest version of the code. My dataset is the nuscenes v1.0-mini dataset, and I have not made any modifications to the code or the dataset. Here are the details of my setup:

Dataset: nuscenes v1.0-mini Configuration: transfusion_nusc_pillar_L.py Hardware: single 3090ti GPU I have also tried training the model for more epochs (up to 50 epochs), but the mAP does not seem to improve significantly.

Could someone please help me diagnose the issue or suggest possible solutions? Thank you!

Feel free to modify the issue text as needed, and provide more details if you think it would be helpful. Good luck with reproducing the Transfusion results!

[2000lf]

@2460707925 Could you tell me how to prepare the nuscenes v1.0-mini dataset

[2460707925]

@2460707925 Could you tell me how to prepare the nuscenes v1.0-mini dataset

juse use mmdet3d create_dataset , set version== v1.0-mini

[docckey]

@2460707925 Could you tell me if you solved the problem of poor training results on the mini dataset.