Datasets:
MANUS-DexYCB Annotation Format
Each WebDataset sample contains a .json metadata member plus the available multimodal files for that sample.
Per-Sample JSON Fields
{
"manus_sample_id": "dexycb_test_000001_processed_000001",
"config": "dexycb",
"split": "test",
"source_dataset": "DexYCB",
"source_version": "DexYCB",
"source_split": "test",
"source_image_id": "000001_processed_000001",
"source_license_id": "cc-by-nc-4.0",
"source_license_name": "Creative Commons Attribution-NonCommercial 4.0 International",
"contains_source_pixels": true,
"processing_pipeline_version": "manus-dexycb-v1.0",
"available_files": ["full_rgb", "full_depth", "hand_rgb", "hand_depth", "vertices", "mesh_front"]
}
Modalities
full_rgb.png: original full-scene RGB image converted to PNG.full_depth.png: full-scene depth map.hand_rgb.png: hand-region RGB crop.hand_depth.png: hand-region depth crop.hand_bbox.json: hand-region bounding box for the RGB crop.hand_depth_bbox.json: hand-region bounding box for the depth crop.vertices.json: local hand mesh vertices and merged bounding-box metadata.mesh_front.png,mesh_back.png,mesh_left.png,mesh_right.png,mesh_top.png,mesh_bottom.png: six canonical rendered mesh views.six_view_bbox_*.json: bounding boxes for canonical rendered mesh views.extra_view_image_primary.png,extra_view_image_complementary.png: selected primary/complementary projection views.extra_view_bbox_primary.json,extra_view_bbox_complementary.json: bounding boxes for selected projection views.
Not every future MANUS subset will expose the same modality set; check available_files and the manifest for each sample.