mzhobro/taco_dataset

TACO: Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding

Dataset Versions

[1] Pre-released Version

Dataset links:

• OneDrive: https://1drv.ms/f/s!Ap-t7dLl7BFUfmNkrHubnoo8LCs?e=1h0Xhe
• BaiduNetDisk: https://pan.baidu.com/s/1gANrhzdUyvsUGXcDB4xMfQ?pwd=kg7j

Dataset Contents:

• 244 high-quality motions sequences spanning 137 <tool, action, object> triplets
• 206 High-resolution object models (10K~100K faces per object mesh)
• Hand-object pose and mesh annotations
• Egocentric RGB-D videos
• 8 allocentric RGB videos

[2] Version 1

Dataset link (Dropbox): https://www.dropbox.com/scl/fo/8w7xir110nbcnq8uo1845/AOaHUxGEcR0sWvfmZRQQk9g?rlkey=xnhajvn71ua5i23w75la1nidx&st=9t8ofde7&dl=0

Dataset link (for a backup, Dropbox): https://www.dropbox.com/scl/fo/6wux06w26exuqt004eg1a/AM4Ia7pK_b0DURAVyxpHLuY?rlkey=e76q06hyj9yqbahhipmf5ij1o&st=c30zhh8s&dl=0

Dataset Contents:

• 2317 motions sequences spanning 151 <tool, action, object> triplets
• 206 High-resolution object models (10K~100K faces per object mesh)
• Hand-object pose and mesh annotations
• Egocentric RGB-D videos
• 12 allocentric RGB videos
• Camera parameters
• Automatic Hand-object 2D segmentations
• Automatic marker-removed images

Resized Version (512x376)

A pre-resized version of this dataset is available at mzhobro/taco_dataset_resized, with all allocentric videos and segmentation masks downscaled to 512x376. This provides ~25x faster data loading and 4.4x less disk space — recommended for training.

	Original (this repo)	Resized	Speedup
Single sample load	3.91 s	158 ms	25x
Throughput (4 workers)	0.52 samp/s	12.2 samp/s	23x
Disk size	2.2 TB	495 GB	4.4x smaller
Allocentric videos	809 GB	4.3 GB	188x smaller

Full profiling details: tools/taco_analysis/profile_comparison.md

Archive Contents

Archive	Contents
Marker_Removed_Allocentric_RGB_Videos.zip	12 camera MP4s per sequence (marker-removed)
Allocentric_RGB_Videos.zip	Original allocentric videos (before marker removal)
2D_Segmentation.zip	Per-camera segmentation masks (750x1024 npy)
Hand_Poses.zip	MANO params (pkl) + pre-computed 3D joints (npy)
Hand_Poses_3D.zip	3D joints only — hand_joints.npy per sequence (T, 2, 21, 3) float32, ~160 MB
Object_Poses.zip	Object 6DoF transforms (npy)
Egocentric_RGB_Videos.zip	Egocentric RGB videos
Egocentric_Depth_Videos.zip.*	Egocentric depth videos (split archive)
object_models_released.zip	206 high-res object meshes
mano_v1_2.zip	MANO hand model files

Tip: If you only need 3D hand joints (not raw MANO parameters), download Hand_Poses_3D.zip (160 MB) instead of Hand_Poses.zip (26 GB).

Downloading the Dataset

The dataset is hosted on HuggingFace. The Allocentric_RGB_Videos/ folder (~484 GB) contains the original recordings before marker removal and is not required for the analysis scripts — those use Marker_Removed_Allocentric_RGB_Videos/ instead. You can exclude either folder to save bandwidth.

Option A: CLI

# Full download (excluding original allocentric videos)
huggingface-cli download mzhobro/taco_dataset \
    --repo-type dataset \
    --exclude "Allocentric_RGB_Videos/*" \
    --local-dir taco_dataset

# Full download (excluding marker-removed videos)
huggingface-cli download mzhobro/taco_dataset \
    --repo-type dataset \
    --exclude "Marker_Removed_Allocentric_RGB_Videos/*" \
    --local-dir taco_dataset

Option B: Python API

from huggingface_hub import snapshot_download

# Full download (excluding original allocentric videos)
snapshot_download(
    "mzhobro/taco_dataset",
    repo_type="dataset",
    ignore_patterns=["Allocentric_RGB_Videos/*"],
    local_dir="taco_dataset",
)

# Full download (excluding marker-removed videos)
snapshot_download(
    "mzhobro/taco_dataset",
    repo_type="dataset",
    ignore_patterns=["Marker_Removed_Allocentric_RGB_Videos/*"],
    local_dir="taco_dataset",
)

After downloading, reassemble split archives and extract:

cd taco_dataset
# Reassemble split files (e.g. Egocentric_Depth_Videos.zip.*.part)
./reassemble.sh
# Extract all zip archives
for z in *.zip; do unzip -qn "$z"; done

Reproducing Analysis Plots

The taco_analysis/ directory contains scripts to generate dataset statistics and visualizations. MANO hand models are provided in mano_v1_2/.

Setup

cd taco_dataset
python -m venv .venv
source .venv/bin/activate
pip install numpy pandas matplotlib seaborn opencv-python imageio Pillow trimesh scipy chumpy

Running the scripts

All scripts auto-detect paths relative to the repository root. Plots are saved under taco_analysis/plots/.

# Dataset summary statistics (11 plots)
python taco_analysis/analyze_taco.py

# 3D camera + object mesh visualization
python taco_analysis/visualize_taco_3d_scene.py

# Top-down / side camera layout
python taco_analysis/visualize_taco_cameras_topdown.py

# Epipolar line grids
python taco_analysis/visualize_taco_epipolar.py

# Object mesh + hand skeleton overlay on allocentric views
python taco_analysis/render_mesh_overlay.py

# Loading performance profile
python tools/taco_analysis/profile_dataset.py --root . --output profile.md

Each script accepts --help for full options (e.g. --n-sequences, --seed, --csv, --root).