byungjun-kim's picture
Release the 48-sample Real-World Dynamic evaluation set
b9e6da6 verified
|
Raw History Blame Contribute Delete
3.2 kB
metadata
license: cc-by-nc-4.0
pretty_name: DWM Real-World Dynamic Evaluation Set
size_categories:
  - n<1K
tags:
  - video
  - egocentric
  - hand-object-interaction
  - world-model
  - project-aria

DWM Real-World Dynamic Evaluation Set

This is the Real-World Dynamic evaluation split of Dexterous World Models (CVPR 2026): 48 egocentric hand–object interaction samples captured with Project Aria glasses under a moving camera.

📄 Paper | 🌐 Project Page | 💻 Code

Contents

Each sample has four aligned files that share the same id:

Folder Description
realworld/videos/ Ground-truth interaction video.
realworld/videos_static/ Static-scene video. The operator explores the scene before acting, a 3D Gaussian scene is reconstructed from those pre-action frames, and it is rendered along the camera trajectory recorded during the interaction.
realworld/videos_hands/ Hand-mesh rendering along the same camera trajectory.
realworld/prompts_rewrite/ Text prompt used for evaluation.

All videos are 720×480, 49 frames, 8 fps, H.264.

realworld_selected_48.txt lists the 48 samples used in the paper. Sample ids follow the original capture numbering, so they are not contiguous.

realworld_selected_48.txt
realworld/
├── videos/00001.mp4 ...
├── videos_static/00001.mp4 ...
├── videos_hands/00001.mp4 ...
└── prompts_rewrite/00001.txt ...

Five of these samples (00013, 00017, 00039, 00043, 00055) are also shipped in the GitHub repo under examples/realworld/realworld_dynamic/; the files are identical.

Usage

Download:

hf download byungjun-kim/DWM-RealWorld-Dynamic --repo-type dataset --local-dir /path/to/dwm_realworld

The layout matches what the DWM inference scripts expect, so the paper evaluation can be run directly from the code repo:

# WAN
bash training/wan/examples/infer_realworld.sh \
  --data_root /path/to/dwm_realworld \
  --dataset_file /path/to/dwm_realworld/realworld_selected_48.txt \
  --checkpoint_path /path/to/DWM-Wan2.1-Fun-14b-LoRA \
  --experiment_config /path/to/DWM-Wan2.1-Fun-14b-LoRA/dwm_wan_14b_lora.yaml

# CogVideoX
bash training/cogvideox/examples/infer_realworld.sh \
  --data_root /path/to/dwm_realworld \
  --dataset_file /path/to/dwm_realworld/realworld_selected_48.txt \
  --checkpoint_path /path/to/DWM-CogVideoX-Fun-5b-LoRA \
  --experiment_config /path/to/DWM-CogVideoX-Fun-5b-LoRA/dwm_cogvideox_5b_lora.yaml

Checkpoints: DWM-Wan2.1-Fun-14b-LoRA, DWM-CogVideoX-Fun-5b-LoRA. See the code repo for checkpoints and the full set of options.

Citation

@inproceedings{kim2026dwm,
  title={Dexterous World Models},
  author={Kim, Byungjun and Kim, Taeksoo and Lee, Junyoung and Joo, Hanbyul},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}