Towards Robust Unsupervised Attention Prediction in Autonomous Driving

Mengshi Qi^1*, Xiaoyang Bi¹, Xianlin Zhang¹, Huadong Ma¹

¹Beijing University of Posts and Telecommunications

This repository is an extension of our ICCV conference paper, "Unsupervised Self-Driving Attention Prediction via Uncertainty Mining and Knowledge Embedding." We propose a robust unsupervised framework that eliminates the need for expensive traffic labels. By leveraging an Uncertainty Mining Branch and a Domain-Specific Prior Enhancement Block, our method bridges the gap between natural and driving scenes. Furthermore, we introduce RoboMixup and the DriverAttention-C benchmark to address the challenges of corruption and central bias in real-world autonomous driving.

• 2025-12-24 🚀 We released DriverAttention-C, a comprehensive benchmark with 126k+ frames for robustness evaluation.
• 2025-01-15 📝 Our extended work is available on arXiv.
• 2023-08-07 🎉 Original work accepted by ICCV 2023! Check the iccv branch for the conference code.

To systematically evaluate robustness, we introduce DriverAttention-C, comprising over 126k frames across synthetic and real-world scenarios. It features 49k+ manually re-annotated frames to ensure ground truth validity under adverse conditions.

The datasets and ground truth labels can be downloaded via:

• Synthetic subsets: Images/Camera Effects | Adverse Weather
• Real-world (Snow): Images | Ground Truth

Extensive experiments demonstrate that our unsupervised method matches or surpasses state-of-the-art fully supervised approaches, reducing corruption degradation by 7.2% and mitigating central bias by 11.2% in terms of KLD.

# Corruption Robustness Training 
python train_robo_cor.py --name exp_name --data-path path/to/data --topK 8 --mix_dir temp_dir

# Mitigating Central Bias Training
python train_longtail.py --name rcpreg --data-path path/to/data --batch-size 4

# Calculate KLD and CC
python test_cor.py --data-path path/to/data --save_model model_name

Note: For SIM, AUC-Borji, AUC-Judd, NSS, please follow the implementation provided in SaliencyMamba metrics.

We demonstrate the importance of attention prediction in autonomous driving decision-making.

1. Prepare data following BDD-OIA.
2. Train the decision model utilizing attention ROIs:

python train_decision.py --name test_ --atten_model {infer_dir} --data-path path/to/data

We would like to thank the authors of SaliencyMamba for their contribution to the community. Part of evaluation metrics code is integrated from SaliencyMamba metrics.

@article{qi2025towards,
  title={Towards Robust Unsupervised Attention Prediction in Autonomous Driving},
  author={Qi, Mengshi and Bi, Xiaoyang and Ma, Huadong},
  journal={arXiv preprint arXiv:2501.15045},
  year={2025}
}


@inproceedings{zhu2023unsupervised,
  title={Unsupervised self-driving attention prediction via uncertainty mining and knowledge embedding},
  author={Zhu, Pengfei and Qi, Mengshi and Li, Xia and Li, Weijian and Ma, Huadong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={8558--8568},
  year={2023}
}