A preprint dated 20 August 2026 reports a computer-vision system called HandMvNet that reconstructs 3D hand joints and the mesh vertices describing the hand’s shape from multiple RGB camera views. The authors report lower relative joint and vertex errors than comparison methods across the evaluated datasets, with camera-independent variants superior in most cases.
The method is calibration-free by design but can optionally incorporate camera parameters. It moves through pre-fusion, fusion and post-fusion stages, using multi-view cross-attention to combine information before producing the 3D joints and mesh vertices.
What the benchmark tables show
The main scores are MPJPErel and MPVPErel: mean position errors for predicted hand joints and mesh vertices, measured in millimetres after root-wrist alignment. The evaluation also includes Procrustes-aligned joint and vertex errors, plus area-under-the-curve scores based on the percentage of correct keypoints.
On the DexYCB-MV test set, the main HandMvNet variant reported 6.73 mm for MPJPErel and 7.19 mm for MPVPErel. Its camera-free version reported 7.03 mm and 7.38 mm, respectively.
On HO3D-MV, the main variant reported 21.43 mm for joint error and 20.17 mm for vertex error. The HR camera-free variant reported lower figures of 20.40 mm and 19.33 mm.
On MVHand, the main model reported 2.07 mm and 7.57 mm, while the HR camera-free version reported 1.77 mm and 7.46 mm. The MediaPipe-DLT comparison listed substantially higher figures of 17.24 mm and 18.42 mm.
Speed and design choices
A figure-based comparison reports that HandMvNet achieved the highest frames-per-second result across the tested camera-view configurations. The report does not give exact FPS values or variability, so it shows a ranking rather than the size of the speed gap.
Internal ablations reported the strongest metric combination with full sinusoidal-plus-joint-plus-crop positional encoding: MPJPErel 6.73, PAJ 4.08 and AUCJ 0.67, compared with 7.69, 4.40 and 0.63 for sinusoidal encoding alone. The same three figures were reported for configurations using five fusion layers and eight camera views.
How far the evidence goes
The datasets used different designs. DexYCB-MV included 10 subjects and eight views per subject; its official S0 split excluded left-hand samples and contained 25,387 training, 1,412 validation and 4,951 test multi-view samples. HO3D-MV was built from seven selected sequences with complete observations from five cameras, using 9,087 training and 2,706 test samples.
MVHand included four subjects and four camera views per subject. Its 21,200 frames were divided into 15,417 training, 1,927 validation and 3,856 test frames.
The paper reports weaker performance on the smaller HO3D-MV dataset and notes that implicitly learning 3D geometry demands substantial data. That is the authors’ interpretation of the benchmark pattern, not an independently tested causal explanation.
The report gives no confidence intervals, statistical significance tests or repeated-run variability estimates. The camera-free variants also were not best on every dataset or metric, despite the overall claim of strong benchmark performance.
These results therefore describe computational reconstruction performance on the named benchmarks. They do not by themselves establish improved human-computer interaction, augmented or virtual reality, robotics or clinical outcomes, and the main scores are aligned relative measures rather than a separate absolute-pose accuracy result.
Paper data and sources
Original title: HandMvNet: Real-Time 3D Hand Pose Estimation Using Multi-View Cross-Attention Fusion
Authors: Muhammad Asad Ali, Nadia Robertini, Didier Stricker
Journal/Repository: Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP (2025), pp. 555-562
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: 10.5220/0013107300003912
Original paper · Full text