FocalComm: Hard Instance-Aware Multi-Agent Perception (WACV 2026)

Pretrained checkpoints for FocalComm, a collaborative perception framework that exchanges hard-instance-oriented features among connected agents via progressive Hard Instance Mining (HIM) and Query-guided Adaptive Feature Fusion (QAFF).

Checkpoints

File	Dataset	Config	AP@0.3/AP@0.5
focalcomm_v2xreal/net_epoch50.pth	V2X-Real (vehicle-centric)	FocalComm 3-stage HIM, epoch 50	Car 91.6/88.4, Pedestrian 53.2/26.1, Truck 50.4/47.2, Overall 65.1/53.9
focalcomm_dairv2x/net_epoch50.pth	DAIR-V2X	FocalComm, epoch 50	Vehicle 73.3/66.4

Each checkpoint directory includes the exact training config.yaml. The DAIR-V2X checkpoint reproduces the paper result exactly; the V2X-Real checkpoint is the final epoch of the paper's training run (within ~2 mAP of the paper's Table 1 row).

Usage

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

python focalcomm/tools/inference.py --model_dir checkpoints/focalcomm_v2xreal --fusion_method intermediate

See the GitHub repository for installation and dataset preparation.

Citation

@inproceedings{shenkut2026focalcomm,
  title={FocalComm: Hard Instance-Aware Multi-Agent Perception},
  author={Shenkut, Dereje and Bhagavatula, Vijayakumar},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2026}
}