MIT CORL 2021

纯视觉BEV方案transformer网络3D检测

paper：[2110.06922] DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries

code：GitHub - WangYueFt/detr3d

1. DNN提图像特征，FPN提多尺度特征

2. pts_bbox_head Detr3DHead

   1. transformer Detr3DTransformer

      1. Detr3DHead
             __init__
                 self.query_embedding = nn.Embedding(self.num_query, self.embed_dims * 2)
          
             forward
                 query_embeds = self.query_embedding.weight
                 hs, init_reference, inter_references = self.transformer(
                     mlvl_feats,
                     query_embeds,
                     reg_branches=self.reg_branches if self.with_box_refine else None,  # noqa:E501
                     img_metas=img_metas,
                 )
          
         Detr3DTransformer
             __init__
                 self.embed_dims = self.decoder.embed_dims
                 self.reference_points = nn.Linear(self.embed_dims, 3)
          
             forward(self, mlvl_feats, query_embed, reg_branches=None, **kwargs):
                 query_pos, query = torch.split(query_embed, self.embed_dims , dim=1)
                 query_pos = query_pos.unsqueeze(0).expand(bs, -1, -1)
          
                 reference_points = self.reference_points(query_pos).sigmoid()

      2. Detr3DCrossAtten

      3. MultiheadAttention

   2. bbox_coder NMSFreeCoder

   3. loss_cls FocalLoss