FoodChallenge

# Food3D-Fresh (FoodChallenge) Dataset · v0.2 > Bosch × Monash (Suzhou) Food Challenge Phase 1 dataset. > Specification: Food3D-Fresh Dataset Spec **v0.7** | Task definition **TASK_DEFINITION v0.6** | Evaluation protocol **BENCHMARK_PROTOCOL v0.6** (bundled with the repository under `foodchallenge_dataset/`). ## Subsets | Subset | Contents | Source | Volume GT protocol | |------|------|------|--------------| | `food3d-fresh-pilot/` | In-house monocular multi-view pilot (pear / baby cabbage / pumpkin, session 20260630; Hikvision CA020-10UC, 1624×1240) | in-house capture | `gt_protocol="direct"` (graduated-cylinder water displacement); mesh GT pending Bosch scanning | | `food3d-fresh-mf3d/` | **Curated MetaFood3D** extension subset (watertight mesh + point cloud + volume/weight GT, categories mapped to this dataset's grouping) | [MetaFood3D](https://lorenz.ecn.purdue.edu/~food3d/) (subject to its original license) | `gt_protocol="mesh_primary"` (divergence-theorem integration over the scanned mesh; nutrition-table volume/weight used for cross-validation) — **Disclosure: this subset has no water-displacement data**, consistent with the mesh_primary precedent in Spec §3.2.4 | ## Directory layout (Spec §5.1) ``` {subset}/ ├── metadata/{dataset_info.json, categories.csv, samples.csv, splits.json, calibration/} └── {group}/{category}/{category}_{NNN}/ ├── sample_info.json # category_l1 / provenance / difficulty_tags / ... └── T0/ ├── rgb_industrial.png # canonical monocular competition input (pilot; mf3d: rendered / RGBD frames backfilled in v2) ├── mesh.obj (+ .mtl/texture) # geometry GT (mm, watertight) ├── pointcloud.ply ├── visible_mask.png / amodal_mask.png # GT masks (intentionally provided inputs) ├── pose/T_cam_mesh.json # pose GT (output of the GT production chain) ├── gt_derived/{rendered_depth,normal,nocs,object_xyz}.exr ├── gt.json # volume{primary_cm3, gt_protocol, gt_confidence, ...} + weight └── scan_meta.json # session / modality status (missing modalities marked pending) ``` Note: an additional reference-object monocular view `rgb_industrial_ref.png` (Blender render with a coplanar checkerboard card, 20 mm squares, beside the food; true size recorded in `gt.json:reference_object`) is 🔄 rendering in progress — it enables the S-IR / S-R settings on the mf3d subset. ## Tasks (see `foodchallenge_dataset/TASK_DEFINITION.md` for details) | Task | Input | Output | Primary metrics | |------|------|------|--------| | **T1** (main leaderboard) | Single RGB + GT mask, 4 settings S-IR/S-I/S-R/S-0 (intrinsics × reference object) | metric mesh / point cloud (mm) + `volume_cm3` | Geometric CD-L2 + volume MARE (rigid SE(3), **scale fitting prohibited**) | | **T1-C** | RGB (±mask, C1/C2) | normalized mesh (longest edge = 1) | normalized Chamfer (Sim(3)) | | **T-D** (primary, pose-free) | N×RGB + per-view K (no poses; D1–D4) | metric mesh + volume + per-view camera poses | Same as T1, plus camera pose RRE/RTE | | **T-D-P** (posed diagnostic) | Same as T-D, plus rig poses (`rig_poses.json`) | metric mesh + volume | Same as T1; separate leaderboard (upper-bound reference) | | **T-E1/E3/E4/E5** | RGB (E4 additionally receives GT mesh) | depth / normal / `T_cam_mesh` / NOCS | AbsRel / angular error / ADD(-S) / L1 | | Track 3 freshness | RGB @ T | grade 1–5 | Acc / MAE (Phase 2) | ## GT production vs. benchmark input isolation (Spec §9.0) Benchmark models **must not access**: GT volume, GT mesh, scanner point clouds, `T_cam_mesh`, Orbbec depth, or any model outputs used in GT production. GT masks are the only intentionally provided derived input. `foodchallenge_dataset/tools/validate_dataset.py --check-input-package` verifies that released input packages contain no leakage. ## Usage ```bash pip install modelscope python -c "from modelscope.msdatasets import MsDataset; \ MsDataset.load('zhiyiJ6666/FoodChallenge')" # Evaluation (the judge implementation is bundled with this repository) python foodchallenge_dataset/eval/evaluate.py \ --pred preds --gt food3d-fresh-pilot --task t1 --setting s0 --out report.json ``` ## Versions | Version | Date | Changes | |------|------|------| | v0.2 | 2026-08 | Schema upgraded to Spec v0.6 (nested volume + provenance); added curated MetaFood3D subset; bundled task-definition / evaluation-code snapshots | | v0.1-pilot | 2026-06-30 | First real capture session (3 samples, multi-view + water-displacement volume GT) |