DADA-seg: Multimodal Accident Scene Segmentation Benchmark

[IROS 2021 & IEEE T-ITS 2021] Official Dataset Repository for the DADA-seg accident scene segmentation benchmark.

• GitHub Repository: https://github.com/InSAI-Lab/ISSAFE
• Model Checkpoints: InSAI-Lab/ISSAFE
• Collection: ISSAFE & EDCNet (IROS 2021 & T-ITS 2021)
• Paper (IROS 2021): ISSAFE: Improving Semantic Segmentation in Accidents by Fusing Event-based Data
• Paper (IEEE T-ITS 2021): Exploring Event-Driven Dynamic Context for Accident Scene Segmentation

---

Dataset Description

Ensuring traffic safety requires robust environmental perception even under severe accident scenarios. Most critical traffic accident scenes involve extreme dynamics, high-speed motions, vehicle rollovers, collisions, and motion blur, causing conventional frame-based RGB segmentation models to degrade drastically.

DADA-seg is a pixel-wise annotated multimodal benchmark selected from DADA-2000. It provides paired:

• RGB image frames (anchor frames and preceding frames)
• High-temporal event camera volume representations (.npz files with temporal bins)
• Fine-grained pixel-level semantic ground truth annotations

The validation set contains 313 carefully verified and annotated accident scenes across challenging traffic conditions.

File Structure

The complete dataset archive DADA-seg.tar.gz (~448 MB) unpacks into the following standard directory structure:

DADA_seg/
├── event/
│   └── val/              # Event volume tensors (.npz)
├── gtFine/
│   └── val/              # Pixel-level semantic annotations (19 Cityscapes-compatible classes)
└── leftImg8bit/
    └── val/              # Corresponding RGB image frames

In addition to DADA-seg.tar.gz, modular archives are also provided:

• dada_val_img_label.tar.gz (~47 MB): Contains leftImg8bit/val and gtFine/val.
• dada_val_event.tar.gz (~402 MB): Contains event/val volume tensors.
• dada_val_label_only.zip (~2.8 MB): Contains only ground truth semantic masks (gtFine/val).

Benchmark Results

Evaluation on DADA-seg demonstrates that incorporating event-driven dynamic context brings +8.2% mIoU gain over RGB-only baselines and outperforms >20 competitive segmentation architectures under accident scenarios.

Method	Modality	Backbone	DADA-seg mIoU (%)	Gain over RGB
ResNet Baseline	RGB	ResNet-18	29.8%	-
ISSAFE (IROS 2021)	RGB + Event	ResNet-18	34.5%	+4.7%
EDCNet (T-ITS 2021)	RGB + Event ($B=2$)	ResNet-18	38.0%	+8.2%

Quick Start & Download

Using the Hugging Face CLI:

# Download complete dataset archive
hf download InSAI-Lab/DADA-seg DADA-seg.tar.gz --type dataset --local-dir ./dataset

# Extract into dataset directory
cd ./dataset
tar -zxvf DADA-seg.tar.gz

Using Python (huggingface_hub):

from huggingface_hub import hf_hub_download

dataset_path = hf_hub_download(
    repo_id="InSAI-Lab/DADA-seg",
    repo_type="dataset",
    filename="DADA-seg.tar.gz",
    local_dir="./dataset"
)

Citation

@inproceedings{zhang2021issafe,
  author={Zhang, Jiaming and Yang, Kailun and Stiefelhagen, Rainer},
  booktitle={2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={ISSAFE: Improving Semantic Segmentation in Accidents by Fusing Event-based Data}, 
  year={2021},
  pages={1132-1139},
  doi={10.1109/IROS51168.2021.9636109}
}

@article{zhang2021edcnet,
  author={Zhang, Jiaming and Yang, Kailun and Stiefelhagen, Rainer},
  journal={IEEE Transactions on Intelligent Transportation Systems}, 
  title={Exploring Event-Driven Dynamic Context for Accident Scene Segmentation}, 
  year={2021},
  pages={1-17},
  doi={10.1109/TITS.2021.3134828}
}

License

This dataset is released under the Apache-2.0 License.