edges listlengths 674 1.35k | id int64 0 799 | metadata dict | nodes listlengths 678 1.35k | occlusion dict |
|---|---|---|---|---|
[{"label":"attached_to","source":"peduncle_d696_s0_a0_v0_1","target":"stem_d696_s0"},{"label":"attac(...TRUNCATED) | 696 | {"collision_stats":{"attempts":600,"failure_pct":60.8333333333,"failures":365,"overlap_avg_pct":1095(...TRUNCATED) | [{"centroid":[0.5657018173,0.9839373149,1.0342522929],"color":[0.8000000119,0.8000000119,0.800000011(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"leaf_d667_s0_a0_v0_1_l0","target":"stem_d667_s0"},{"label":"attach(...TRUNCATED) | 667 | {"collision_stats":{"attempts":600,"failure_pct":57.6666666667,"failures":346,"overlap_avg_pct":694.(...TRUNCATED) | [{"centroid":[0.5662327461000001,0.9848324390000001,0.9564730467],"color":[0.8000000119,0.8000000119(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"leaf_d63_s0_a0_v0_1_l0","target":"stem_d63_s0"},{"label":"attached(...TRUNCATED) | 63 | {"collision_stats":{"attempts":750,"failure_pct":61.2,"failures":459,"overlap_avg_pct":1017.58008894(...TRUNCATED) | [{"centroid":[0.5632397955,0.9809271057000001,1.1115132283],"color":[0.8000000119,0.8000000119,0.800(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"peduncle_d533_s0_a0_v0_1","target":"stem_d533_s0"},{"label":"attac(...TRUNCATED) | 533 | {"collision_stats":{"attempts":750,"failure_pct":59.7333333333,"failures":448,"overlap_avg_pct":867.(...TRUNCATED) | [{"centroid":[0.5681730498,0.9871179545000001,0.7891268343000001],"color":[0.8000000119,0.8000000119(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"peduncle_d66_s0_a0_v0_1","target":"stem_d66_s0"},{"label":"attache(...TRUNCATED) | 66 | {"collision_stats":{"attempts":600,"failure_pct":59.3333333333,"failures":356,"overlap_avg_pct":1139(...TRUNCATED) | [{"centroid":[0.5629649286,0.9805368397,1.1076380553],"color":[0.8000000119,0.8000000119,0.800000011(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"peduncle_d621_s0_a0_v0_1","target":"stem_d621_s0"},{"label":"attac(...TRUNCATED) | 621 | {"collision_stats":{"attempts":750,"failure_pct":63.7333333333,"failures":478,"overlap_avg_pct":720.(...TRUNCATED) | [{"centroid":[0.5679878946,0.9866026159000001,0.8742407521000001],"color":[0.8000000119,0.8000000119(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"leaf_d346_s0_a0_v0_1_l0","target":"stem_d346_s0"},{"label":"attach(...TRUNCATED) | 346 | {"collision_stats":{"attempts":900,"failure_pct":61.6666666667,"failures":555,"overlap_avg_pct":690.(...TRUNCATED) | [{"centroid":[0.5624755189,0.9798880298,1.1177915813000001],"color":[0.8000000119,0.8000000119,0.800(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"leaf_d490_s0_a0_v0_1_l0","target":"stem_d490_s0"},{"label":"attach(...TRUNCATED) | 490 | {"collision_stats":{"attempts":900,"failure_pct":62.1111111111,"failures":559,"overlap_avg_pct":860.(...TRUNCATED) | [{"centroid":[0.5678561544,0.9869509782,0.848975565],"color":[0.8000000119,0.8000000119,0.8000000119(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"peduncle_d760_s0_a0_v0_1","target":"stem_d760_s0"},{"label":"attac(...TRUNCATED) | 760 | {"collision_stats":{"attempts":600,"failure_pct":63.8333333333,"failures":383,"overlap_avg_pct":867.(...TRUNCATED) | [{"centroid":[0.5681377059,0.9875743544000001,0.7702940135],"color":[0.8000000119,0.8000000119,0.800(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
[{"label":"attached_to","source":"leaf_d456_s0_a0_v0_1_l0","target":"stem_d456_s0"},{"label":"attach(...TRUNCATED) | 456 | {"collision_stats":{"attempts":900,"failure_pct":59.2222222222,"failures":533,"overlap_avg_pct":1071(...TRUNCATED) | [{"centroid":[0.5679118353,0.9869045878,0.8443157125],"color":[0.8000000119,0.8000000119,0.800000011(...TRUNCATED) | {"dirs":[[1.0,0.0,0.0],[-1.0,0.0,0.0],[0.0,1.0,0.0],[0.0,-1.0,0.0],[0.0,0.0,1.0],[0.0,0.0,-1.0],[-0.(...TRUNCATED) |
SG-DOR Pepper Scene Graphs
SG-DOR Pepper Scene Graphs is a synthetic dataset of 800 procedurally generated pepper-plant scenes. Each scene combines an attributed plant-organ graph, an instance-labeled 3D point cloud voxelized at 4 mm, and direction-conditioned leaf-fruit occlusion targets. It was prepared from the training inputs used by the SG-DOR project.
The default Hugging Face configuration is graph_only, intended for graph machine learning without a robotics stack. The full point-cloud and occlusion data are stored as WebDataset-compatible tar shards under data/full/.
Dataset contents
Each scene contains:
graph.json: nodes for stems, leaves, shoots, peduncles, and fruits; labeled structural edges; oriented bounding boxes; centroids; orientations; colors; generation metadata; and the original OBB-based directional occlusion annotations.pointcloud.npz: 4 mm point cloud with coordinates, normals, colors, semantic labels, and instance labels.occlusion.npz: point-cloud z-buffer targets over 18 fruit-local directions, including fruit union occlusion and per leaf-fruit potential, exclusive, and three-layer depth-rank targets.
There are 640 training scenes and 160 validation scenes. The split exactly reproduces the SG-DOR Oracle loader: shuffle scene indices 0..799 with NumPy seed 42, then take the first 80% as training data. No independent test split is supplied.
Aggregate counts, ranges, and label frequencies are recorded in dataset_info.json. Exact scene membership and SHA-256 hashes are recorded under metadata/.
Loading the graph-only configuration
After publication, replace the placeholder repository ID below:
from datasets import load_dataset
graphs = load_dataset("<namespace>/sg-dor-pepper-scene-graphs", "graph_only")
scene = graphs["train"][0]
print(len(scene["nodes"]), len(scene["edges"]))
For a local package:
from datasets import load_dataset
graphs = load_dataset(
"json",
data_files={
"train": "data/graph_only/train.jsonl.gz",
"validation": "data/graph_only/validation.jsonl.gz",
},
)
Loading full scenes
The tar members belonging to a scene share a key, for example:
scene_00000.graph.json
scene_00000.pointcloud.npz
scene_00000.occlusion.npz
They can be streamed with WebDataset or read with Python's tarfile module. A dependency-light local example is included:
python scripts/load_scene.py data/full/train-00000-of-00016.tar scene_00696
See docs/SCHEMA.md for field definitions and label mappings.
Intended uses
The dataset supports research on attributed and heterogeneous graph learning, node and edge prediction, geometric graph neural networks, 3D point-cloud learning, occluder ranking, directional relation prediction, multi-task learning, and uncertainty estimation. Graph-only users can ignore the point-cloud and occlusion archives.
Limitations and out-of-scope uses
- All scenes are synthetic. Performance on these scenes does not establish generalization to real plants, other crops, field conditions, sensors, or robot platforms.
- Plant geometry, organ placement, colors, topology, and occlusion statistics reflect the procedural generator and its asset library.
- The release has only one deterministic 80/20 train-validation split and no held-out test set.
- OBB-based occlusion embedded in graph JSON and point-cloud z-buffer occlusion in sidecar NPZ files are different annotation sources. They should not be treated as interchangeable.
- Semantic and instance labels are generated from simulation rather than human annotation.
- The data should not be used to claim biological realism, agricultural safety, or autonomous manipulation reliability without separate validation.
The dataset contains no people or personal data.
Generation and preprocessing
Pepper plants were procedurally assembled in Blender/BlenderProc. Plant organs were exported as graph nodes and structural edges. Mesh surfaces were converted to labeled point clouds and voxelized at 4 mm. Z-buffer occlusion sidecars were then computed from the 4 mm point clouds using a 4 mm pixel size, 4 mm depth tolerance, a 0.2 m leaf-fruit candidate radius, 18 fruit-local viewing directions, and three stored depth ranks.
The packaged release intentionally excludes the redundant denser point clouds and standalone metadata JSON files. The graph JSON already contains the generation metadata used for each scene. No training label or input used by the confirmed 4 mm SG-DOR Oracle runs is omitted.
Reproducibility and integrity
metadata/scenes.csv: split membership, loader order, source sizes, and per-component SHA-256 hashes.metadata/splits.json: exact ordered scene indices for both splits.metadata/shards.json: shard membership summaries, sizes, and hashes.SHA256SUMS: hashes for all package files.scripts/build_release.py: deterministic release builder.scripts/validate_release.py: structural and checksum validation.
License
The dataset and accompanying documentation are released under the MIT License. See LICENSE.
Citation
@inproceedings{menon2026sgdor,
title={{SG-DOR}: Learning Scene Graphs with Direction-Conditioned Occlusion Reasoning for Pepper Plants},
author={Menon, Rohit and Mueller-Goldingen, Niklas and Pan, Sicong and Chenchani, Gokul Krishna and Bennewitz, Maren},
booktitle={Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2026},
note={Accepted}
}
Preprint: https://arxiv.org/abs/2603.06512
Upload
Create a dataset repository and upload this directory from its parent folder:
hf repo create <namespace>/sg-dor-pepper-scene-graphs --repo-type dataset
hf upload <namespace>/sg-dor-pepper-scene-graphs sg_dor_huggingface . --repo-type dataset
Uploading is intentionally not performed by the release builder.
- Downloads last month
- -