CompetitorFormer

CompetitorFormer: Mitigating Query Conflicts for 3D Instance Segmentation via Competitive Strategy (CVPR 2026)

Duanchu Wang, Junjie Yang, Haoran Gong, Jing Liu, Di Wang

[Paper] [Code]

Official checkpoints for CompetitorFormer, a competitive-strategy-based transformer that mitigates query conflicts for 3D instance segmentation.

Results & Models

ScanNet++ (official 50-scene val)

• Model: CompetitorFormer on ScanNet++ official 50-scene val
• Checkpoint: competitorformer_scannetpp.pth (num_query=500)
• AP / AP₅₀ / AP₂₅ = 0.339 / 0.485 / 0.581

mkdir -p checkpoints
wget https://huggingface.co/WangDuanchu/CompetitorFormer/resolve/main/competitorformer_scannetpp.pth \
  -O checkpoints/competitorformer_scannetpp.pth

or:

huggingface-cli download WangDuanchu/CompetitorFormer competitorformer_scannetpp.pth --local-dir checkpoints

python tools/train.py configs/scannetpp/competitorformer_scannetpp.yaml \
  --work_dir exps/competitorformer_scannetpp \
  --eval_only \
  --resume checkpoints/competitorformer_scannetpp.pth

Inference knobs (do not reuse ScanNet v2 values): num_query=500, topk_insts=1300, num_class=84.

ScanNet v2

• Checkpoint: competitorformer_scannet.pth (num_query=400)

wget https://huggingface.co/WangDuanchu/CompetitorFormer/resolve/main/competitorformer_scannet.pth \
  -O checkpoints/competitorformer_scannet.pth

or:

huggingface-cli download WangDuanchu/CompetitorFormer competitorformer_scannet.pth --local-dir checkpoints

See the GitHub README for installation, data preparation, training, and evaluation.

Citation

@InProceedings{Wang_2026_CVPR,
  title     = {CompetitorFormer: Mitigating Query Conflicts for 3D Instance Segmentation via Competitive Strategy},
  author    = {Wang, Duanchu and Yang, Junjie and Gong, Haoran and Liu, Jing and Wang, Di},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {34724-34733},
  year      = {2026},
}

License

This project is released under the MIT License.