File size: 1,716 Bytes
dac9e9c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
---
license: mit
tags:
- collaborative-perception
- v2x
- 3d-object-detection
- autonomous-driving
- lidar
---

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

Pretrained checkpoints for [FocalComm](https://github.com/scdrand23/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

```bash
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](https://github.com/scdrand23/FocalComm) for installation and dataset preparation.

## Citation

```bibtex
@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}
}
```