关于reference_points和predicted_3dbbox

[kingofstu]

您好，首先很高兴你们能有这样伟大的工作！其次，我想提两个问题：
1、在BevFormer的decoder中：
query_pos, query = torch.split( object_query_embed, self.embed_dims, dim=1) # [900,256], [900,256]
query_pos = query_pos.unsqueeze(0).expand(bs, -1, -1)
query = query.unsqueeze(0).expand(bs, -1, -1) # [B,900,256]
reference_points = self.reference_points(query_pos) # linear [B,900,3]
reference_points = reference_points.sigmoid()
这个reference_points是否可以理解为anchor的坐标？ 它是全局坐标系下的坐标还是bev自车坐标系下的坐标呢？

2、当使用reg_branches时，预测的10个量分别表示什么，看上去像是6层xy的偏移加上reference_points的xy，tmp[
..., 4:5]+reference_points的z表示什么呢？predicted_3dbbox的坐标是全局坐标系下的坐标还是bev自车坐标系下的坐标？
if reg_branches is not None:
tmp = reg_brancheslid # 6个回归的头，每个都是2层mlp [B,900,10]
assert reference_points.shape[-1] == 3
new_reference_points = torch.zeros_like(reference_points)
new_reference_points[..., :2] = tmp[
..., :2] + inverse_sigmoid(reference_points[..., :2])
new_reference_points[..., 2:3] = tmp[
..., 4:5] + inverse_sigmoid(reference_points[..., 2:3])
new_reference_points = new_reference_points.sigmoid()
reference_points = new_reference_points.detach()