[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
Not using square point cloud range get very low match ious #64
Hi Xuyang, thanks for your work firstly!
I am trying to train this model(backbone with PointPillars) with different point cloud range and voxel size.
My point cloud range is [-70, -42, -5, 98, 42, 3], and my voxel size is [0.15, 0.15, 8]
When I only modify the config, I get a very low match ious during training, alwarys around 0.00xx.
I notice that in transfusion_head.py, function create_2d_grid, I will get batch_y and batch_x with 280x140, but my heatmap size should be 140x280, so I change the meshgird to [[0, y_size -1, y_size], [0, x_size -1, x_size]] and get 2 140x280 tensors, then my match ious increase to 0.2x but still low.
Do you have any idea with this problem? thx a lot
Hi Xuyang, thanks for your work firstly! I am trying to train this model(backbone with PointPillars) with different point cloud range and voxel size. My point cloud range is [-70, -42, -5, 98, 42, 3], and my voxel size is [0.15, 0.15, 8] When I only modify the config, I get a very low match ious during training, alwarys around 0.00xx. I notice that in transfusion_head.py, function create_2d_grid, I will get batch_y and batch_x with 280x140, but my heatmap size should be 140x280, so I change the meshgird to [[0, y_size -1, y_size], [0, x_size -1, x_size]] and get 2 140x280 tensors, then my match ious increase to 0.2x but still low. Do you have any idea with this problem? thx a lot