megvii-research / Far3D

[AAAI2024] Far3D: Expanding the Horizon for Surround-view 3D Object Detection
Other
138 stars 11 forks source link

About nuScenes #9

Closed xiazhongyv closed 9 months ago

xiazhongyv commented 9 months ago

感谢您杰出的工作! 想参考下您在nuscenes数据集上的config,不知是否可以分享,我的邮箱是 xiazhongyu@pku.edu.cn,非常感谢!

以及我在生成argoverse 2 info文件时,会有数行如下关于数据集图像缺失的提示: No corresponding ring image at 315969904359876000: 27.5 > 25.0 ms No corresponding ring image at 315969904359876000: 40.1 > 25.0 ms No corresponding ring image at 315969904359876000: 32.6 > 25.0 ms No corresponding ring image at 315966060559935000: 27.5 > 25.0 ms No corresponding ring image at 315969614859771000: 27.7 > 25.0 ms No corresponding ring image at 315969614859771000: 32.7 > 25.0 ms ...... 想请教下这个是av2数据集的正常现象吗?

xhjiang-pixel commented 9 months ago

1)config已发;2)是正常的,AV2官方设置的图像同步阈值是25ms,我们当时处理数据时扩大到了100ms,不过train集还是有缺少图片的情况,训练时会跳过这些样本。

xiazhongyv commented 9 months ago

非常感谢! 还有一个问题,当我把batchsize从1调到2时,会报错: File "/home/xiazhongyu/bevperception/mmdet3d/models/dense_heads/farhead.py", line 816, in build_query2d_proposal
raise NotImplementedError 您有遇到过这个情况吗?

xhjiang-pixel commented 9 months ago

抱歉,因为目前我们采用threshold_score去选取adaptive_query时还没有做多batch_size适配,如果采用topk方式采样的话可以应用多batch_size

Yuning-Shi commented 3 months ago

@xiazhongyv @xhjiang-pixel 您好,可以分享一份训练nuscenes时使用的config文件吗,我的邮箱是 "19ru601@ms.dendai.ac.jp"。非常感谢!!!