XuyangBai / TransFusion

[PyTorch] Official implementation of CVPR2022 paper "TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection with Transformers". https://arxiv.org/abs/2203.11496
Apache License 2.0
613 stars 76 forks source link

About coordinate system refactoring #99

Open lilkeker opened 1 year ago

lilkeker commented 1 year ago

Hi, I encountered the problem of mAP and NDS indicators being lower than your results when reproducing TransFusion_Lidar. I noticed [this],(https://github.com/XuyangBai/TransFusion#:~:text=Please%20refer%20to%20data_preparation,size%20(l%2C%20w).) The reason is that I used a higher version of mmdet3d (1.00rc4) for training and eval, mmdet3d in v1.0.0 The coordinate system has been reconstructed , Does the correct TransFusion_Lidar result only can be achieved under the lower version of mmdet3d? I now want to implement the correct TransFusion_Lidar on mmdet3d (1.00rc4), but I don’t know which parts to modify. I would be very grateful if you could give me some advice and guidance! my result: e0ed8441e5faca5d16e87d15d9a8ae2

lilkeker commented 1 year ago

Hi, I encountered the problem of mAP and NDS indicators being lower than your results when reproducing TransFusion_Lidar. I noticed [this],(https://github.com/XuyangBai/TransFusion#:~:text=Please%20refer%20to%20data_preparation,size%20(l%2C%20w).) The reason is that I used a higher version of mmdet3d (1.00rc4) for training and eval, mmdet3d in v1.0.0 The coordinate system has been reconstructed , Does the correct TransFusion_Lidar result only can be achieved under the lower version of mmdet3d? I now want to implement the correct TransFusion_Lidar on mmdet3d (1.00rc4), but I don’t know which parts to modify. I would be very grateful if you could give me some advice and guidance! my result: e0ed8441e5faca5d16e87d15d9a8ae2

I have solved the mAOE and mASE problems caused by the version, but I still cannot reach the TransFusion_Lidar index mentioned in the paper by adopting the fading strategy. My result mAP: 0.5608 NDS: 0.6438 , I noticed that many people have encountered the same problem. When I checked my Log, I found that my loss may not be normal: 2023-06-02 23:50:03,576 - mmdet - INFO - Iter [120000/120105] lr: 1.534e-08, eta: 0:01:10, time: 0.653, data_time: 0.008, memory: 5365, lossheatmap: 0.7224, layer-1_losscls: 0.1174, **layer-1_loss_bbox: 0.8705, matched_ious: 0.5250, loss: 1.7103, grad_norm: 1.0566 2023-06-02 23:50:36,541 - mmdet - INFO - Iter [120050/120105] lr: 1.149e-08, eta: 0:00:36, time: 0.659, data_time: 0.008, memory: 5365, lossheatmap: 0.7201, layer-1_loss_cls: 0.1169, layer_-1_loss_bbox: 0.8421, matched_ious: 0.5269, loss: 1.6791**, grad_norm: 1.0716 How to solve this?