MuViSeg checkpoints

Trained heads for MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors (ACCV 2026).

• Code: https://github.com/MuViSeg/muviseg-codeRelease
• Project page: https://muviseg.github.io
• Paper: https://arxiv.org/abs/2607.17938

Given two or more images of the same scene plus class-agnostic segmentation masks for each, MuViSeg predicts which segments correspond to the same physical object. A frozen 3D foundation-model backbone (MASt3R or VGGT) supplies dense descriptors, masked average pooling gives one descriptor per segment, and a lightweight trainable head (LightGlue-style attention with a DoubleSoftmax matcher) predicts the assignment, with a dustbin for unmatched segments.

These are head-only checkpoints

The frozen backbone is stripped and reloaded separately at inference time, so each file is a few MB rather than a few GB. Each holds the trainable head plus epoch, global_step, best_val_ma and the validation metrics. They cannot resume training — that needs the optimizer state of a full checkpoint.

The backbones are not hosted here; they belong to their upstream authors: VGGT-1B from facebook/VGGT-1B, and MASt3R from the MASt3R release.

file	model	trainable params	size
segmast3r_lg_v2/best.pth	SegMASt3R + LG v2	~0.8 M	3.3 MB
segvggt_dpt/v3-001/best.pth	SegVGGT-DPT, pairwise	~10.7 M	50 MB
segvggt_dpt/joint-001/best.pth	SegVGGT-DPT, joint attention over N frames	~11.0 M	49 MB
segvggt/best.pth	SegVGGT single-layer (ablation)	~0.8 M	5.5 MB
segvggt/step_0140000.pth	SegVGGT single-layer, step 140k	~0.8 M	5.5 MB

The single-layer row reported in the paper uses segvggt/best.pth on Replica and segvggt/step_0140000.pth on Virtual KITTI 2, which is why both are published.

Results

Overall AUPRC / R@1 / R@5, on the frozen benchmark pair lists shipped with the code.

model	Replica	Virtual KITTI 2
SegMASt3R + LG v2	84.5 / 77.5 / 93.6	81.2 / 72.7 / 92.8
SegVGGT-DPT (pairwise)	79.7 / 74.1 / 90.1	51.9 / 42.9 / 67.6
SegVGGT-DPT Joint (N=4)	81.6 / 76.5 / 90.6	52.8 / 43.2 / 66.8
SegVGGT single-layer	79.2 / 73.3 / 89.6	51.5 / 42.1 / 66.3

MASt3R + LG v2 dominates at small viewpoint change; joint multi-frame attention narrows the gap at large rotations. The code repository's docs/reproduction.md states which rows are exactly reproducible and which are not — read it before comparing against the paper.

Usage

git clone https://github.com/MuViSeg/muviseg-codeRelease.git
cd muviseg-codeRelease
uv sync
bash setup_third_party.sh
uv run python scripts/download_checkpoints.py --all --verify

That places every file where the shipped configs expect it. To fetch one directly:

from huggingface_hub import hf_hub_download

path = hf_hub_download("MuViSeg/muviseg", "segvggt_dpt/v3-001/best.pth")

License

MIT, as the code. Third-party components (MASt3R, VGGT, SegMASt3R, LightGlue, RoMa, SAM 2, FastSAM) remain under their own licenses.

Citation

@inproceedings{muviseg2026,
  title     = {MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors},
  author    = {Fatykhoph, Denis and Akhtyamov, Timur and Pakulev, Konstantin
               and Devchich, German and Ferrer, Gonzalo},
  booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
  year      = {2026}
}