JVC
Main checkpoint for JVC: Joint Vision Cross-Attention for Fine-Grained Badminton Stroke Recognition (PDF). Navneeth Dhamotharan and Bin Han, ECCV 2026 Workshop: Human Motion-Informed World Models and Socially Intelligent Action, 2026.
The released weight is the joint-vision cross-attention model.
A 16-frame hit clip goes through two encoders: R(2+1)D Conv3D for RGB patches, and a four-stream SkateFormer for MediaPipe joints. Joint tokens cross-attend to visual patches, a divided space–time transformer mixes the clip, and contact-weighted pooling feeds multitask heads. The paper metric is 9-way stroke_type.
Val stroke_type accuracy | 80.61% (epoch 18) |
| Split | Video-level 80/20, seed 42. No clip from a validation video is in training. |
| Dataset | FineBadminton-20K |
| Frames | 16, span_linspace over the hit span, 224×224, ImageNet normalization |
| Skeleton | MediaPipe, 33 joints × (x, y, z), four-stream (joint, bone, joint-motion, bone-motion) |
| Paper | JVC: Joint Vision Cross-Attention for Fine-Grained Badminton Stroke Recognition |
| Demo | isocourt.fit |
Stroke classes, in logit order: Serve, Clear, Smash, Drop, Drive, Net_Shot, Lob, Defensive_Shot, Other.
The checkpoint also emits logits for technique, placement, position, intent, and quality. stroke_type is the number reported above.
Try the demo
Upload a clip at isocourt.fit.
Weights
from huggingface_hub import hf_hub_download
path = hf_hub_download("navneethdg/JVC", "jvc.pth")
frames is (batch, 16, 3, 224, 224) float RGB after ImageNet normalization
(mean = (0.485, 0.456, 0.406), std = (0.229, 0.224, 0.225)).
pose is (batch, 16, 33, 3) in the same layout as the training MediaPipe cache.
Heads return logits, not probabilities.
Files
| File | Role |
|---|---|
jvc.pth | Checkpoint. Constructor metadata sits beside the weights. |
config.json | Architecture, label names, and metric. |
What this weight is
Vision backbone r2plus1d_18 (conv3d), embed dim 128, 2 cross-attention layers, 4 divided space–time blocks, four-stream skeleton. Shuttle features are off.
No-cross-attention JVC ablations are separate runs and are not this file.
Citation
@inproceedings{
dhamotharan2026jvc,
title={{JVC}: Joint Vision Cross-Attention for Fine-Grained Badminton Stroke Recognition},
author={Navneeth Dhamotharan and Bin Han},
booktitle={ECCV26 Workshop: Human Motion-Informed World Models and Socially Intelligent Action},
year={2026},
url={https://openreview.net/forum?id=XJEhcXfwEe}
}