TactoHand
English
TactoHand predicts contact, proximity, normalized force and visibility for both hands from RGB video. This release uses a DINOv3 ViT-L backbone, 256-pixel inputs and 32-frame windows sampled at 10 Hz.
Download tactohand.ckpt and its matching tactohand.yaml together. The YAML is a
portable inference configuration; the checkpoint also retains training state.
RGB-only inference does not require MANO assets or prepared datasets. Force
outputs are normalized predictions, not forces in newtons.
TactoHand original contributions are licensed under MIT. This checkpoint includes DINOv3-derived backbone weights, which remain subject to the DINOv3 License. The backbone was initialized from timm's DINOv3 ViT-L weights.
简体中文
TactoHand 从 RGB 视频预测双手的接触、距离、归一化力和可见性。 本次发布使用 DINOv3 ViT-L 骨干,输入分辨率为 256,以 10 Hz 采样 32 帧窗口。
请同时下载 tactohand.ckpt 和配套的 tactohand.yaml。YAML 是可独立使用的推理配置;
checkpoint 还保留训练状态。仅 RGB 推理无需 MANO 文件或预处理数据集。
力输出是归一化预测值,不能直接视为牛顿单位的物理力。
TactoHand 自有部分采用 MIT 许可。checkpoint 中包含的 DINOv3 衍生骨干权重仍须遵守 DINOv3 上游许可。 使用方法见代码仓库。
Model tree for zhenqis123/TactoHand
Base model
timm/vit_large_patch16_dinov3.lvd1689m