wenjingbian/viewstruct_train

ViewStruct training corpus

Training data of ViewStruct: View-Grounded Articulation and Connectivity Inference (Wenjing Bian, Ranran Huang, Andreas Geiger, 2026).

• Paper: https://arxiv.org/abs/2610.16793
• Project page: https://bianwenjing.github.io/viewstruct/
• Code: https://github.com/bianwenjing/viewstruct

This is the exact mixture the released ViewStruct model (LoRA on Qwen3-VL-4B-Instruct) was fine-tuned on. Every row is a camera-native articulation example: the model sees an RGB view, a part-segmentation map and a depth map of the same camera (plus extra rest-pose views or an articulation clip for some rows) and emits rigid bodies and ordered joints in a gravity-aligned object-centred frame.

Contents

JSONL (jsonl/)	source	rows	sampled per epoch
physx_cnv3grav_structure_train.jsonl	PhysX-Mobility objects (1,633 objects, canonical + 7 random cameras)	13,064	100 %
physx_aug2_cnv3grav_structure_train.jsonl	PhysX-Mobility decomposition variants (4,561 variants)	36,488	50 %
physxnet_cnv3grav_structure_train.jsonl	PhysXNet objects (40,000 objects, canonical + 3 random cameras)	160,000	40 %
physx_video_cnv3grav_structure_train.jsonl	PhysX-Mobility articulation clips (16 frames, one moving part)	5,954	100 %
physx_cnv3grav_grscenes_structure_train.jsonl	GRScenes-100 articulated objects (2,502 objects)	10,008	100 %
physx_cnv3grav_grscene_structure_train.jsonl	GRScenes-100 room views with several articulated objects	2,338	100 %

Total 227,852 rows, about 114k examples per epoch with the sampling rates above (2 epochs in the paper).

archive	files	size	what
camnative_v2_image.tar	626,502	3.05 GB	*_view.png, *_seg.png, *_depth.png of every row (the three inputs)
physx_image.tar	2,405	30 MB	extra rest-pose RGB views of PhysX-Mobility objects
physx_image_aug2.tar	2,492	30 MB	same for the decomposition variants
physxnet_image.tar	80,000	0.49 GB	extra rest-pose RGB views of PhysXNet objects
physx_video_v2.tar	5,954	0.45 GB	articulation clips (<oid>_l<link>_k<cam>/video.mp4)

The archives are plain uncompressed tar (PNG/MP4 are already compressed); they contain only the files referenced by the six JSONLs.

Row format

{
  "image": ["datasets/my-dataset/camnative_v2_image/<cat>/<name>_cn0_view.png",
            "datasets/my-dataset/camnative_v2_image/<cat>/<name>_cn0_seg.png",
            "datasets/my-dataset/camnative_v2_image/<cat>/<name>_cn0_depth.png", "..."],
  "video": ["datasets/my-dataset/physx_video_v2/<oid>_l<link>_k<cam>/video.mp4"],   // clip rows only
  "conversations": [{"from": "human", "value": "<image>\n<image>\n<image>\n\nYou are a 3D reconstruction agent. ..."},
                    {"from": "gpt",   "value": "{\"n_joints\": 3, \"base\": [..], \"joints\": [{\"parts\": [..], \"on\": \"base\", \"type\": \"hinge\", \"pivot_px\": [u, v], \"axis\": [x, y, z], \"position\": [x, y, z], \"range\": [lo, hi]}, ...]}"}],
  "camnative": {"target": "v3", "frame": "gravity", "source": "physx", "oid": "...", "k": 0, "s": 0.95, "c_cam": [...], "fov_deg": 40.0, ...}
}

• The prompt lists the 2D box and depth range of every segmented region, the field of view and the joint count, and defines the output frame: gravity-aligned, origin at the silhouette centre / median object depth, unit = the larger silhouette extent. Room views additionally list the objects and their regions.
• The target is a JSON with the static body (base), and one entry per joint: the moving parts, what they are attached to, hinge/slide, the axis, the hinge pivot (pivot_px in 0..1000 image units and position in the object frame) and the motion range (radians or object units). Clip rows name the one moving_part.
• camnative records the frame (s, c_cam), the source dataset and object id and the camera index k (0 = canonical camera), so predictions can be mapped back to world units for scoring.

Image paths are relative to the qwen-vl-finetune/ directory of the code repository (datasets/my-dataset/...), so after extracting the archives into qwen-vl-finetune/datasets/my-dataset/ the training script runs unchanged.

Download and train

git clone https://github.com/bianwenjing/viewstruct.git && cd viewstruct
bash download_viewstruct_from_hf.sh            # ~5 GB, extracts the archives
cd qwen-vl-finetune && bash scripts/sft_qwen3_4b_physx_cnv3grav_grs_lora.sh

or by hand:

from huggingface_hub import snapshot_download
snapshot_download("wenjingbian/viewstruct_train", repo_type="dataset",
                  local_dir="qwen-vl-finetune/datasets/my-dataset")
# then: mv jsonl new_jsonl; for t in *.tar; do tar -xf $t; done   (inside that directory)

Sources and licences

The renders are derived from third-party assets. Use of this corpus is bound by their terms:

• PhysX-Mobility and PhysXNet (Cao et al., PhysX-Anything / PhysXGen): CC BY-NC 4.0.
• GRScenes-100 (GRUtopia, Wang et al.): CC BY-NC-SA 4.0.

The corpus as a whole is therefore released under CC BY-NC-SA 4.0 (non-commercial, share-alike). The prompts, joint annotations and the camera-native conversion are ours.

Citation

@article{bian2026viewstruct,
  title   = {ViewStruct: View-Grounded Articulation and Connectivity Inference},
  author  = {Bian, Wenjing and Huang, Ranran and Geiger, Andreas},
  journal = {arXiv preprint arXiv:2610.16793},
  year    = {2026}
}