Closed StiphyJay closed 1 year ago
Following FSD, fade strategy is used during training process which brings +1.0 L2 mAPH gain on Cyclist. The fade strategy is a common trick in Waymo and Nuscenes, and we will incorporate it in camera ready version if necessary.
Thank you, @chenshi3, for the quick reply. @StiphyJay If you have any questions, please feel free to reach out to us. :)
Haiyang,
Thanks for your reply! It is necessary to report the performance with and without fade strategy. Especially for different categories. Fade strategy usually brings improvement in all categories.
Thanks for your good suggestions. However, considering paper space limitations (8 pages in camera-ready) and additional experiments on other datasets will be added (SOTA performance on NuScenes with a remarkable gain), we will not include this result in the main paper, but we will mention that fade strategy was used. The corresponding results (without the fade) will also be released in this repo. Actually, it's a common trick used in all SOTA nuscenes and Waymo approaches, such as BEVFusion, Transfusion and FSD. Moreover, in our experiments, the fade strategy only worked on cyclists(+1.0). Pedestrians will go up a little(+0.5), and hardly be any improvement in the Veh(+0.1~0.2).
Hope that can ease your concern!
Best regards, Haiyang
Thanks for your good suggestions. However, considering paper space limitations (8 pages in camera-ready) and additional experiments on other datasets will be added (SOTA performance on NuScenes with a remarkable gain), we will not include this result in the main paper, but we will mention that fade strategy was used. The corresponding results (without the fade) will also be released in this repo. Actually, it's a common trick used in all SOTA nuscenes and Waymo approaches, such as BEVFusion, Transfusion and FSD. Moreover, in our experiments, the fade strategy only worked on cyclists(+1.0). Pedestrians will go up a little(+0.5), and hardly be any improvement in the Veh(+0.1~0.2).
Hope that can ease your concern!
Best regards, Haiyang
We will consider your valuable advice. Maybe we can add it to the appendix. Thanks for your good suggestions again, which make our paper more substantial. :)
Best regards, Haiyang
Thanks for your reply!
Best regards!
It seems that the performance of cyclists class in your paper is higher than that of pedestrians and vehicles? But there are more sample sizes for vehicle(4352210) and pedestrian(2037627) categories, while cyclists only have 49518 samples in the waymo training dataset, which doesn't seem reasonable. Did you use the fade strategy during your training process?