Results 9 issues of

你的注释写的也太详细了,我追的女孩子看你的代码都不需要我了,555555~

Hi Author: ` def deform_sampling(self, feat, offset): """ Sampling the feature x according to offset. Args: feat (Tensor): Feature offset (Tensor): Spatial offset for for feature sampliing """ # it...

## 📚 Documentation * Links to the relevant documentation/comment: ` reg_targets[l] = reg_targets[l] / float(self.strides[l])`

documentation

请问c_y, c_x = h * float(stride), w * float(stride)在映射回原图的方法为什么和fcos论文提到的ws+1/2s和hs+1/2s不一样呢?

indexing with dtype torch.uint8 is now deprecated, please use a dtype torch.bool instead. how to avoid this warning when training?

I have a problem when run "c.execute("SELECT id FROM Game")",the error is no such table: Game.I wanna impove this model,can we discuss it together?

hello author: I find that fsdv2 model training occupy less memory, config is configs/fsdv2/fsdv2_nusc_2x.py, and I want use fsdv2 for autolabel, so do you have any suggestions on the model...

作者您好: (1)我将DAL集成到了自己的训练框架(也是基于mmdet3d),实验发现在自己数据集上检测结果确实优于centerpoint很多,但是将图像的backbone从res18变为res34的时候发现指标没有明显增长。是不是图像特征其实不是很重要? (2)从标注的思路上来考虑,回归分支的heading angle和box size是不是也可以考虑用融合特征,而不是单一的点云特征,中心位置只用点云特征。我在自己数据集上发现纯点云检测对于远一点的目标(>70m)的heading angle容易翻转180°,所以想用图像特征来弥补远处点云稀疏造成的heading angle不准确。