StZaBL's picture
Upload README.md with huggingface_hub
ef74e47 verified
metadata
license: mit
task_categories:
  - robotics
  - object-detection
language:
  - en
tags:
  - robotics
  - grasping
  - affordance
  - hand-object-interaction
  - contact-detection
  - YCB
  - DexYCB
  - HO3D
  - FoundationPose
  - MANO
pretty_name: Affordance2Grasp Preprocessed Assets
size_categories:
  - 1G<n<10G

Affordance2Grasp — Preprocessed Assets

Preprocessed object meshes and initialization files for the Affordance2Grasp pipeline.

This repository contains assets you download once before running the pipeline. The pipeline then generates per-dataset intermediate data (depth maps, MANO estimates, poses, contact labels) from the raw videos, which are not included here due to size and licensing constraints.


Contents

Folder Size Description
obj_meshes/ycb/ ~1 GB YCB object meshes + scale.json (metric scale) for all objects used in DexYCB and HO3D v3
obj_recon_input/ycb/ ~50 MB FoundationPose initialization masks (reference images + bounding boxes) for each YCB object
training_fp/dexycb/ ~1 MB Validated contact labels for DexYCB ycb_dex_01 and ycb_dex_02 (sanity-check reference)

Usage

pip install huggingface_hub

python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='StZaBL/Affordance2Grasp-ProcessedData',
    repo_type='dataset',
    local_dir='data_hub/ProcessedData',
)
"

Then follow the pipeline in the main repository README.


Supported Datasets

Both datasets use the same YCB objects — the meshes here cover all required objects.

Dataset Objects covered
DexYCB 20 YCB objects (ycb_dex_01ycb_dex_20)
HO3D v3 003_cracker_box, 004_sugar_box, 006_mustard_bottle, 010_potted_meat_can, 011_banana, 021_bleach_cleanser, 035_power_drill, 052_extra_large_clamp

Raw Data

Raw videos must be downloaded separately from the official sources:


Related

  • Code: stzabl-png/UCB_Project
  • Pipeline: Depth Pro → HaPTIC → FoundationPose → Contact Labels → Affordance Model → Robot Grasp