Marsil123 commited on
Commit
d2a92ff
路
verified 路
1 Parent(s): 527352c

UA-Fit HRNet-W40 weights (network only) + model card

Browse files
Files changed (2) hide show
  1. README.md +75 -0
  2. uafit_w40.safetensors +3 -0
README.md ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ library_name: pytorch
4
+ tags:
5
+ - hand-pose-estimation
6
+ - hand-mesh-recovery
7
+ - multi-view
8
+ - mano
9
+ - 3d-reconstruction
10
+ pipeline_tag: image-to-3d
11
+ ---
12
+
13
+ # UA-Fit (HRNet-W40)
14
+
15
+ **Learn 2D, Solve 3D: UA-Fit, an Uncertainty-Weighted Analytical Solver for Multi-view Hand Mesh Recovery.**
16
+ ECCV 2026 Workshops (HANDS). Marsil Zakour, Constantin Patsch, Martin Piccolrovazzi, Yuankai Wu, Eckehard Steinbach,
17
+ Technical University of Munich.
18
+
19
+ [Project page](https://marsilzakour.github.io/ua-fit/) 路 [Code](https://github.com/MarsilZakour/ua-fit-code) 路
20
+ [Solver (parafit)](https://github.com/MarsilZakour/parafit)
21
+
22
+ The network predicts per-view 2D evidence (a dense joint vote field with a per-vote 2x2 precision), triangulates it
23
+ with its uncertainty, and an analytical Levenberg-Marquardt solver fits a parametric hand to it. The same weights fit
24
+ MANO (the paper), Meta's MHR hand or the UmeTrack hand.
25
+
26
+ ## Files
27
+
28
+ | file | content |
29
+ |---|---|
30
+ | `uafit_w40.safetensors` | network weights (HRNet-W40 backbone, vote-field / uncertainty / shape heads), 189 MB |
31
+
32
+ The file contains only the network. It does **not** contain MANO model data: download MANO from
33
+ https://mano.is.tue.mpg.de/ (its own license) as described in the code repository.
34
+
35
+ ## Usage
36
+
37
+ ```bash
38
+ git clone https://github.com/MarsilZakour/ua-fit-code && cd ua-fit-code
39
+ pip install -r requirements.txt
40
+ python scripts/evaluate.py --checkpoint Marsil123/ua-fit # downloads uafit_w40.safetensors
41
+ python scripts/evaluate.py --checkpoint Marsil123/ua-fit --hand-model mhr
42
+ ```
43
+
44
+ ```python
45
+ from uafit.network import load_uafit_network
46
+ network = load_uafit_network("Marsil123/ua-fit", device="cuda")
47
+ ```
48
+
49
+ ## Results (Table 1, full official test splits, mm)
50
+
51
+ | dataset (views) | MPVPE | MPJPE | PA-MPVPE |
52
+ |---|---|---|---|
53
+ | DexYCB (8) | 7.86 | 7.55 | 4.92 |
54
+ | HO3D (5) | 8.01 | 7.41 | 5.10 |
55
+ | ARCTIC (8) | 7.54 | 6.62 | 5.66 |
56
+ | InterHand2.6M (8) | 8.77 | 7.79 | 6.23 |
57
+ | OakInk (4) | 9.58 | 8.86 | 5.67 |
58
+
59
+ Reproduced with these weights and the released code (`scripts/evaluate.py`).
60
+
61
+ ## Training data and license
62
+
63
+ Trained on multi-view DexYCB, HO3D, ARCTIC, InterHand2.6M, OakInk and FreiHAND. Several of these are released for
64
+ non-commercial research only; the weights are therefore released under CC BY-NC 4.0 for research use.
65
+
66
+ ## Citation
67
+
68
+ ```bibtex
69
+ @inproceedings{zakour2026uafit,
70
+ title = {Learn 2D, Solve 3D: UA-Fit, an Uncertainty-Weighted Analytical Solver for Multi-view Hand Mesh Recovery},
71
+ author = {Zakour, Marsil and Patsch, Constantin and Piccolrovazzi, Martin and Wu, Yuankai and Steinbach, Eckehard},
72
+ booktitle = {European Conference on Computer Vision (ECCV) Workshops -- HANDS},
73
+ year = {2026}
74
+ }
75
+ ```
uafit_w40.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:83f7fe48cdd80fc666a38a124b18255d14866864ac462e6da477a088181939a6
3
+ size 189179648