Explainable Saliency: Articulating Reasoning with Contextual Prioritization

Nuo Chen, Ming Jiang, Qi ZhaoCVPR 2025 · OpenReview

Abstract

Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by building a saliency model that can not only identify what is important in an image, but also explain its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model's attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us.

Introduction

Saliency models—algorithms that predict which parts of an image humans will look at—are essential for applications like autonomous driving, virtual reality, and medical diagnostics. However, most state-of-the-art models are "black boxes"; they provide a heatmap of importance but offer no insight into why a region was chosen.

Existing attempts at "explainable" saliency often fall short, either relying on massive predefined vocabularies of object labels or providing simple descriptions of eye fixations without any underlying logical reasoning. Human attention is a multifaceted process driven by visual cues, semantic understanding, and specific tasks. To truly mimic human gaze, a model must be able to articulate its reasoning process in a way that aligns with human cognition.

The XSal Framework

The XSal (Explainable Saliency) architecture moves beyond "black-box" predictions by providing step-by-step natural language explanations. It consists of three main stages:

  1. Explicit Reasoning: The model utilizes a Vision-Language Model (VLM), specifically MiniCPM-V 2.6, to perform two-stage prompting. It first "plans" the analysis based on the context (task-driven or free-viewing) and then extracts specific semantic proposals and their corresponding rationales.
  2. Visual Grounding: These proposals are fed into ClipSeg, an open-vocabulary segmentation model, to generate spatial masks. These are refined into semantic prototypes that represent the visual grounding of the VLM's reasoning.
  3. Contextual Prioritization: Instead of a dense combination of all features, XSal employs a Progressive Top-K Reduction strategy. It weights each semantic prototype and selects only the most relevant ones, mimicking human selective attention. A transparent linear layer then combines these top prototypes into the final saliency map.
Figure 1: Overview of the XSal framework demonstrating task-driven reasoning and contextual prioritization.

Experimental Evaluation

The authors evaluated XSal on two distinct eye-tracking datasets: AiR (Attention in Reasoning) for task-driven gaze and OSIE (Object and Semantic Images and Eye-tracking) for free-viewing complexity.

Key Results

  • State-of-the-Art Accuracy: XSal outperformed existing models (including SAM and DINet) on the AiR dataset with a Normalized Scanpath Saliency (NSS) of 1.94 and on the OSIE dataset with an NSS of 2.96.
  • Superior Explainability: Using XAI metrics like AOPC and LOdds, XSal proved significantly more sensitive to the removal of important semantic concepts compared to previous models, confirming that its explanations directly drive its predictions.
  • Qualitative Success: The model successfully articulates complex relationships, such as identifying a "fire hydrant" as salient specifically because it is "to the right of the garbage can" as requested by a user query.

Discussion and Future Work

XSal demonstrates that reasoning acts as a powerful prior. By generating a natural language plan before prediction, the heatmap is grounded in logic rather than just low-level pixel features. Furthermore, the power of sparsity is evident; focusing on a few contextually relevant prototypes (Top-K) actually improves performance and interpretability. The use of open-vocabulary models like ClipSeg ensures flexibility, allowing the model to reason about any object the VLM can describe.

However, the model's high selectivity may discard low-saliency regions that humans occasionally glance at, making it "cleaner" but potentially less representative of full human "noise." Additionally, the integration of a full VLM introduces computational overhead. Future work will explore extending this framework to video saliency and "human-in-the-loop" systems for interactive reasoning refinement.

Summary updated

Abstract

Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by building a saliency model that can not only identify what is important in an image, but also explain its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model's attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us.
BibTeX
@inproceedings{chen2025explainable,
  author = {Nuo Chen and Ming Jiang and Qi Zhao},
  title = {{Explainable Saliency: Articulating Reasoning with Contextual Prioritization}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  url = {https://openreview.net/forum?id=vI2t6QnGNN},
}
Figure