ReVQom: Residual Vector Quantization for Communication-Efficient Multi-Agent Perception (ICASSP 2026)

Pretrained checkpoints for ReVQom, a learned feature codec for multi-agent collaborative perception. ReVQom compresses BEV features via a 1x1 bottleneck and multi-stage residual vector quantization, transmitting only per-pixel code indices (6-30 bits per pixel, 273x-1365x compression vs raw features).

Checkpoints

File	Config	Dataset	AP@0.3/AP@0.5
revqom_s_k64_dairv2x/net_epoch30.pth	ReVQom-S: K=64, n_q=3, C_rr=16, EMA 0.8 (18 bpp, 455x)	DAIR-V2X	0.751/0.646

The checkpoint directory includes the exact training config.yaml, and the evaluation matches the paper's experiment records. Additional checkpoints (ReVQom-M, K=256) will be added after re-evaluation.

Usage

pip install -U huggingface_hub
hf download scdrand23/ReVQom --local-dir checkpoints

python revqom/tools/inference.py --model_dir checkpoints/revqom_s_k64_dairv2x --fusion_method intermediate

See the GitHub repository for installation and dataset preparation.

Citation

@inproceedings{shenkut2026revqom,
  title={Residual Vector Quantization for Communication-Efficient Multi-Agent Perception},
  author={Shenkut, Dereje and Kumar, B.V.K. Vijaya},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2026}
}