Lightweight BEV Transformer – Attention & Temperature

• Generates synthetic BEV grids (32×32) with rectangles standing in for:
  • cars (5×3)
  • pedestrians (2×2)
• Trains a small transformer over BEV patches:
  • patch embedding: 4×4 → 8×8 = 64 patches
  • 2 encoder layers, 2 heads, 64-dim embeddings
• Uses object queries + soft-argmax detection head:
  • predicts object centers (x, y in [0, 1])
  • predicts class (car vs pedestrian)
• Exposes internal self-attention maps so you can:
  • overlay them on the BEV grid
  • compare different temperatures in the softmax

The whole thing is CPU-friendly and stays under ~100k parameters.