search

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
619 stars 76 forks source link

Performance of TransFusion-L with 3 decoder layers #21

Open WWW2323 opened 2 years ago

WWW2323 commented 2 years ago

Hi, when I set the num_decoder_layers=3 in the config file, the performance of TransFuion-L drops about 1.5NDS. In the table 6 of paper, 3 decoder layers can bring some improvement, do I need to modify anything of codes to achieve similar results? Thanks~

XuyangBai commented 2 years ago

Hi, you might need to increase the loss_bbox from 0.25 to 3*0.25 because the bounding box loss will be divided by the total number of positive predictions. If 3 decoder layers are used, the total number of positive predictions will be 3 times that for 1 decoder layer.

zwei-lin commented 1 year ago

Hi @WWW2323, I also want to try to increase the number fo decoder layers. Do you get a higher result when increasing the loss_bbox from 0.25 to 3*0.25, as XuyangBai mentioned?