一、背景

基于imp投影的相機視角轉換，對相機的內外參依賴較高，BEV 網格融合固定，可能對小目標不夠敏感；考慮通過transformer的方式進行相機的視角轉換，BEV query 可以自適應關注關鍵區域，提高小目標檢測，transformer 注意力機制，靈活采樣。故通過BEVFormer的demo代碼理解其原理。

二、代碼

import torch
import torch.nn as nn# -------------------------
# 參數
# -------------------------
B, C, H, W = 2, 64, 16, 16        # 攝像頭特征圖
bev_H, bev_W = 8, 8                # BEV 網格
num_cameras = 7
num_classes = 10
num_det_queries = 32                # detection query 數量# -------------------------
# 1. 多攝像頭特征
# -------------------------
camera_feats = [torch.randn(B, C, H*W) for _ in range(num_cameras)]  # B x C x N (N=H*W)
for i in range(num_cameras):camera_feats[i] = camera_feats[i].permute(0, 2, 1)  # B x N x C# -------------------------
# 2. BEV query + Transformer 投影
# -------------------------
num_bev_queries = bev_H * bev_W
bev_queries = nn.Parameter(torch.randn(num_bev_queries, B, C))class BEVProjectionTransformer(nn.Module):def __init__(self, C, num_heads=8):super().__init__()self.attn = nn.MultiheadAttention(embed_dim=C, num_heads=num_heads)def forward(self, bev_queries, camera_feats):"""bev_queries: num_bev_queries x B x Ccamera_feats: list of B x N x C"""# 拼接所有攝像頭特征feats = torch.cat(camera_feats, dim=1)      # B x (num_cameras*N) x Cfeats = feats.permute(1,0,2)               # (num_cameras*N) x B x Cbev_out, _ = self.attn(bev_queries, feats, feats)return bev_outbev_proj_transformer = BEVProjectionTransformer(C)
bev_features = bev_proj_transformer(bev_queries, camera_feats)
bev_features_grid = bev_features.permute(1,0,2).reshape(B, bev_H, bev_W, C)# -------------------------
# 3. Detection query + Transformer
# -------------------------
det_queries = nn.Parameter(torch.randn(num_det_queries, B, C))class DetectionDecoderTransformer(nn.Module):def __init__(self, C, num_heads=8):super().__init__()self.attn = nn.MultiheadAttention(embed_dim=C, num_heads=num_heads)def forward(self, det_queries, bev_features_grid):B, H, W, C = bev_features_grid.shapebev_flat = bev_features_grid.reshape(B, H*W, C).permute(1,0,2)out, _ = self.attn(det_queries, bev_flat, bev_flat)return outdecoder = DetectionDecoderTransformer(C)
det_features = decoder(det_queries, bev_features_grid)# -------------------------
# 4. Detection head
# -------------------------
class SimpleDetectionHead(nn.Module):def __init__(self, C, num_classes):super().__init__()self.cls_head = nn.Linear(C, num_classes)self.bbox_head = nn.Linear(C, 7)def forward(self, det_features):cls_logits = self.cls_head(det_features)bbox_preds = self.bbox_head(det_features)return cls_logits, bbox_predsdetection_head = SimpleDetectionHead(C, num_classes)
cls_logits, bbox_preds = detection_head(det_features)print("類別 logits shape:", cls_logits.shape)     # num_det_queries x B x num_classes
print("3D bbox preds shape:", bbox_preds.shape)   # num_det_queries x B x 7