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
640 stars 77 forks source link

About the number of predictions #61

Closed Teut1 closed 2 years ago

Teut1 commented 2 years ago

Really interesting and exciting work. I was wondering about the number of detections TransFusion predicts. When visualizing my training results I always get full 200 detections (num_proposals). In the visualizations of your paper that does not seem to be the case. Did you modify the score_threshhold in the bbox_coder or am I missing something else? In #48 you also provided the evaluation results for the training without the fade strategy, and it looks like there are also around 155 predictions per sample (your results). If that is the case, do you know why is this not a problem for the mAP metric?