Question about sample topk of foreground points?

[eagle-chase]

if self.training and self.train_cfg.get('disable_pretrain', False) and not self.runtime_info.get('enable_detection', False):
            seg_scores = seg_scores[:, cls_id]
            topks = self.train_cfg.get('disable_pretrain_topks', [100, 100, 100])
            k = min(topks[cls_id], len(seg_scores))
            top_inds = torch.topk(seg_scores, k)[1]
            fg_mask = torch.zeros_like(seg_scores, dtype=torch.bool)
            fg_mask[top_inds] = True

Does anyone know why here we need to filter the points with the top 500 highest score for each category by score? After sampling, each class only have 500 points, Will it lose a lot of information? In my understanding, this is the point that constitutes all virtual voxels. Why is such a small threshold needed? Hope someone can give me some help, I'm very confused on this part, thanks a lot！

[Abyssaledge]

This is only enabled at the very beginning of training. You can refer to this hook for details: https://github.com/tusen-ai/SST/blob/main/mmdet3d/core/hook/fsd_hooks.py#L70C11-L70C11

[eagle-chase]

Thank you！