Explainable Saliency: Articulating Reasoning with Contextual Prioritization

This repository contains the code and resources for the CVPR 2025 paper: Explainable Saliency: Articulating Reasoning with Contextual Prioritization. Our work introduces a novel saliency prediction model that not only identifies important image regions but also explains its reasoning process using a vision-language approach and contextual prioritization.

---

🔍 Overview

Our approach bridges the gap between saliency prediction and interpretability by incorporating explicit reasoning and contextual prioritization mechanisms:

• Explicit Reasoning: Uses a vision-language model to generate semantic proposals and natural language explanations that mirror human attention and problem-solving processes.
• Contextual Prioritization: Dynamically selects the most relevant features to focus on key semantic elements in the image.

Key Features:

1. Saliency Prediction with Reasoning: Produces state-of-the-art saliency maps alongside explanations for why specific regions are prioritized.
2. Human-Like Interpretability: Generates natural language descriptions that explain the model’s decisions in a manner aligned with human reasoning.
3. Dynamic Contextual Focus: Adapts to scene complexity, emphasizing only the most relevant semantic elements.

⚙️ Setup

Prerequisites:

• Python 3.8
• PyTorch 1.9.1
• Install dependencies:

  pip install -r requirements.txt

📂 Dataset Preparation:

The air_data dataset can be downloaded from the following link:
Download air_data Structure the root directory as follows:

Explainable_Saliency/
  |-- data/
      |-- air_data/
      |-- osie_data/
  |-- src/
  |-- bash/
  |-- README.md
  |-- requirements.txt

---

🚀 Training

Train the model on both datasets:

python ./src/AirReasonTrainer.py --mode=train --epoch=10 --lr=4e-4 --batch_size=10 --topk=3 --checkpoint_dir=./workdir

📊 Evaluation

Evaluate the trained model:

python AirReasonTrainer.py --mode=eval --topk=3 --checkpoint_dir=./workdir/your_checkpoints_path

---

✍️ Citation

If you find our work helpful in your research, please consider citing it:

@InProceedings{Chen_2025_CVPR,
    author    = {Chen, Nuo and Jiang, Ming and Zhao, Qi},
    title     = {Explainable Saliency: Articulating Reasoning with Contextual Prioritization},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {9601-9610}
}