XuyangBai
commented
2 years ago Hi @AndyYuan96 Thanks for your interest in our work. The fade strategy indeed has a large effect on the final performance by reducing many FP, and I never hide this fact. And similar with your experiment, we have tried TransFusionL_VoxelNet without the fade strategy and find it achieves 61.4 mAP & 68.1 NDS on the nuScenes benchmark, you can check it here. And this performance already surpasses CenterPoint (58.1 mAP & 67.1 NDS) with similar settings.
Personally, Transformer is good because 1). the transformer-based detection head with bipartite matching loss alleviates the need for NMS post-processing, and makes the detection system end-to-end and fast. 2) the transformer architecture is naturally compatible with our soft-association idea where the multi-modal fusion is conducted using an attention mechanism to dynamically learn what and where to fuse. 3). the attention provides the change to leverage long-term dependency and a large receptive field, which could be helpful for detection or other downstream tasks.
AndyYuan96
Hi， xuyang，I reproduce the result of transfusion-l-pillar model, as it did have improvement compared with center point-pillar, I just wonder where the improvement come from? So I did some experiment. I find when I use some training config which is used by transfusion(like use 10 sweep, setting dataset config's valid_mask=false, gtaug fade strategy) to train centerpoint-pillar, centerpoint's performance also boost, and interestingly, centerpoint's nds is higher than transfusion-l-pillar more than 1point, and centerpoint's map is lower than transfusion only 0.1 point. what's more, For all the experiment, I only do once. So it's very interesting that whether transformer is indeed better 。