• 🔥Update
• ⚡Proposed Model
• 📖Datasets
• 🛠️Deployment
  • [1] Environment
  • [2] Run train
  • [3] Run test
  • [4] Run visualization
  • [5] Sota One/Few-shot
• 🚀Demo
• ⭐️Cite

• 2025/07/16: We provide two handy scripts.

  • modify_yaml.py: One-click to modify dataset path. script
  • collect_metrics.py: One-click to collect metrics data. script

• 2025/06/20: All the code and models are completed.

  • How to configure: command & script
  • How to train: command
  • How to evaluate: command
  • How to visualize: command

• 2025/04/19: We collect existing meta-learning methods for driver saliency prediction.

  • Environment configuration
    • metadriver_ori: PFENet (TPAMI'22), BAM (CVPR'22). Details
    • metadriver_mmcv: HDMNet (CVPR'23), AMNet (NIPS'23), AENet (ECCV'24). Details
  • How to use: command & script

• 2025/02/15: We collect driving accident data with data categories, divided into 4 folds by category.

  • DADA: 52 categories. It repartitioned into DADA-52i.
  • PSAD: 4 categories. It repartitioned into PSAD-4i.

• 2025/01/11: We propose a model in order to learn to learn driver attention in driving accident scenarios.

we propose a framework that combines a one-shot learning strategy with scene-level semantic alignment, enabling the model to achieve human-like attentional perception in previously unseen scenarios. (named MetaDriver)

1. A Novel One-Shot Learning Framework for Driver Saliency Prediction. We propose MetaDriver, inspired by how drivers generalize from limited examples. It filters salient regions from support masks to guide attention learning and uses semantic alignment to improve cross scene generalization.
2. A Novel One-Shot Learning Framework for Driver Saliency Prediction. To address noisy attention labels and limited semantic overlap, SMF removes irrelevant background while SSA aligns salient features between support and query pairs, significantly boosting our model’s performance.
3. Extensive Evaluation on DADA-52i and PSAD-4i with Strong Generalization and SOTA Results. It exceeds 10 baselines on both datasets and different backbones, achieving up to +34.3% CC, +45.2% SIM and –31.2% KLD gains under oracle-support. These results highlight its robustness and backbone-agnostic design.

【note】 For all datasets we will provide our download link with the official link. Please choose according to your needs.

DADA-52i: This dataset we will upload in BaiduYun (please wait). Official web in link.

PSAD-4i: This dataset we will upload in BaiduYun (please wait). Official web in link.

MetaDADA/
├── 1/
│   ├── 001/ 
│   │   ├── images/   
│   │   └── maps/  
│   ├── 002/
│   └── ...
├── 2/
│   ├── 001/ 
│   │   ├── images/   
│   │   └── maps/  
│   ├── 002/
│   └── ...
└── ...

MetaPSAD/
├── images/
│   ├── 2/ 
│   │   ├── 0004/   
│   │   └── ...  
│   ├── 3/
│   └── ...
└── maps/
    ├── 2/ 
    │   ├── 0004/   
    │   └── .../  
    ├── 3/
    └── ...

• Git clone this repository

git clone https://github.com/zhao-chunyu/MetaDriver.git
cd MetaDriver

• Create conda environment

conda create -n MetaDriver python=3.11
conda activate MetaDriver

• Install PyTorch 2.5.1+cu121

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121

• Install requirement

pip install -r utils/requirements.txt

Before training, scripts you might use: { modify_yaml.py: One-click to modify dataset path. Details }

If you wish to train with our model, please use the command below.

sh scripts/train_MetaDriver.sh [datasaet] [split#] [backbone] [gpu]

dataset: metadada, metapsad.

#: 0, 1, 2, 3.

backbone: resnet50, vgg.

gpu: 0 or other.

​ Example:

sh scripts/train_MetaDriver.sh metadada split0 resnet50 0

We calculate the predicted values and then use python for the prediction.

sh scripts/test_MetaDriver.sh [datasaet] [split#] [backbone] [gpu]

​ Example:

sh scripts/test_MetaDriver.sh metadada split0 resnet50 0

After testing, scripts you might use: { collect_metrics.py: One-click to collect metrics data. Details }

If you want to visualize all the data of a certain dataset directly, you can use the following command.

sh scripts/visual_MetaDriver.sh [datasaet] [split#] [backbone] [gpu]

​ Example:

sh scripts/visual_MetaDriver.sh metadada split0 resnet50 0

If you want to run other comparison models, we have supported 5 SOTA models. Details

Then, you can run the following command.

sh scripts/[train_*.sh] [dataset] [split#] [backbone] [gpu]

*: PFENet, BAM, HDMNet, AMNet, AENet.

dataset: metadada, metapsad.

#: 0, 1, 2, 3.

backbone: resnet50, vgg.

gpu: 0 or other.

​ Example:

sh scripts/train_PFENet.sh metadada split0 resnet50 0

Model result testing and model result visualization are similar to our MetaDriver model.

• Compare with SOTA Models

• Case of Unseen Accident Scenario

If you find this repository useful, please use the following BibTeX entry for citation and give us a star⭐.

waiting accepted