[PyTorch] Official implementation of CVPR2022 paper "TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection with Transformers". https://arxiv.org/abs/2203.11496
Thank you for open-sourcing your work. I was wondering about the image-guided query initialization, it seems the image collapse feature only used to generate heatmap, but not used for the following query initialization, the center is remapped to lidar-only bev feature other than image-fused bev feature for center query embedding initialization. Another question is that when adopting the image-guided query initialization, it seems reserving the lidar-only bev feature to generate heatmap and average the dense_heatmap and the dense_heatmap_img, i wonder why not only use the dense_heatmap_img?
The related source codes are as follows, looking forward for your reply ~
https://github.com/XuyangBai/TransFusion/blob/73c596f7bd3460c17cbcc58dd9bcc5a0896774a8/mmdet3d/models/dense_heads/transfusion_head.py#L813-L871
Thank you for open-sourcing your work. I was wondering about the image-guided query initialization, it seems the image collapse feature only used to generate heatmap, but not used for the following query initialization, the center is remapped to lidar-only bev feature other than image-fused bev feature for center query embedding initialization. Another question is that when adopting the image-guided query initialization, it seems reserving the lidar-only bev feature to generate heatmap and average the dense_heatmap and the dense_heatmap_img, i wonder why not only use the dense_heatmap_img? The related source codes are as follows, looking forward for your reply ~ https://github.com/XuyangBai/TransFusion/blob/73c596f7bd3460c17cbcc58dd9bcc5a0896774a8/mmdet3d/models/dense_heads/transfusion_head.py#L813-L871