konpat/huron-sacson-hdf5

Huron Sacson HDF5 Dataset

Real-world robot navigation trajectories from ETH Zurich, stored in efficient HDF5 format for faster I/O and easier sharing.

Dataset Overview

The Huron Sacson dataset contains real-world robot navigation trajectories collected from December 2022 to February 2023. This HDF5 version provides a single-file format that significantly improves I/O performance compared to loading individual JPEG files.

Key Statistics

• Trajectories: 2,955 sequences
• Samples: ~165,000 valid samples (with 4-4-4 frame configuration)
• Format: HDF5 with JPEG-encoded frames
• Image Size: 224x224 pixels (configurable)
• Date Range: December 2022 - February 2023
• Dataset Size: ~950MB (HDF5 file)

Frame Configuration

The dataset uses a flexible input/gap/output structure:

• Input frames: Configurable (default: 4)
• Gap: Configurable (default: 4)
• Output frames: Configurable (default: 4)

This allows for various temporal modeling tasks including:

• Video prediction (predicting future frames)
• Representation learning (learning from temporal sequences)
• Contrastive learning (comparing input and target sequences)

Dataset Structure

huron_hdf5/
├── sacson.h5              # Main HDF5 file with train/test splits
├── data_splits/
│   ├── train/            # Training split trajectory lists
│   └── test/             # Test split trajectory lists
├── data_config.yaml      # Dataset configuration
└── README.md             # This file

HDF5 File Structure

The sacson.h5 file contains:

• /train/ group: Training trajectories
  • Each trajectory contains a frames dataset with JPEG-encoded frame bytes
  • Trajectory attributes stored in group metadata
• /test/ group: Test trajectories
  • Same structure as training set

Usage

Download from Hugging Face

# Download the dataset
huggingface-cli download <username>/huron-sacson-hdf5 \
    --repo-type dataset \
    --local-dir ./datasets/huron_hdf5/

Loading the Dataset

Using PyTorch

from huron_hdf5_dataset import HuronHDF5Dataset

# Create dataset
dataset = HuronHDF5Dataset(
    hdf5_path='datasets/huron_hdf5/sacson.h5',
    split='train',
    dataset_name='sacson',
    input_frames=4,
    output_frames=4,
    gap=4,
    image_size=224,
    transform=None,  # or processor name like 'videomae-base'
    normalize=True
)

# Access samples
sample = dataset[0]
input_frames = sample['input']      # Shape: (4, 3, 224, 224)
target_frames = sample['target']    # Shape: (4, 3, 224, 224)
video_key = sample['video_key']     # Unique identifier

Using PyTorch Lightning

The dataset is integrated with the unified video data module:

data:
  dataset_name: "huron_sacson"
  huron_config:
    hdf5_path: "datasets/huron_hdf5/sacson.h5"
    use_hdf5: true
  input_frames: 4
  output_frames: 4
  gap: 4

Dataset Interface

The HuronHDF5Dataset class implements the BaseVideoDataset interface, providing:

• __len__(): Get dataset size
• __getitem__(idx): Get sample by index
• find_index_by_video_key(key): Find index by video key
• get_sample_metadata(idx): Get sample metadata
• get_dataset_metadata(): Get overall dataset metadata

Benefits of HDF5 Format

Compared to the original JPEG file format:

• Faster I/O: Single file access vs 241K+ individual file operations
• Easier Sharing: One file to download and manage
• Better Performance: Reduced filesystem overhead
• Compatible: Drop-in replacement for JPEG-based loader

Data Splits

The dataset includes pre-computed train/test splits:

• Training split: Located in data_splits/train/
• Test split: Located in data_splits/test/

Each split contains trajectory name lists and optional pre-computed indices for faster loading.

Configuration

The data_config.yaml file contains dataset-specific parameters:

• Action statistics (min/max values)
• Metric waypoint spacing
• Other trajectory metadata

Citation

If you use this dataset in your research, please cite the original Huron dataset:

@article{huron2023,
  title={Huron: Real-world Robot Navigation Dataset},
  author={ETH Zurich},
  year={2023},
  note={Dataset collected December 2022 - February 2023}
}

License

Please refer to the original Huron dataset license for usage terms.

Support

For issues or questions:

• Check the main project README: README.md
• Review dataset loading code: huron_hdf5_dataset.py
• Test dataset loading: python scripts/test_huron_hdf5.py

Related Files

• Dataset Loader: huron_hdf5_dataset.py - Main dataset class
• Conversion Script: scripts/convert_huron_to_hdf5.py - Convert from JPEG to HDF5
• Test Script: scripts/test_huron_hdf5.py - Verify dataset integrity
• Original Dataset: datasets/huron/sacson/ - Original JPEG format