Research map / Vision-language and multimodal / Human and embodied modeling
Position: Do Not Explain Vision Models Without Context
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that visual XAI methods ignore spatial context and proposes shifting explanations from "where" to "how" objects relate, with illustrative XAI failures on KITTI and dSprites.
Problem
Popular post-hoc visual explanation methods highlight input regions but fail to capture spatial relationships (distance, inside/outside, order, orientation) that determine labels in critical domains such as surveillance, autonomous driving, and healthcare.
Approach
Argues explanations should move from "where" (heatmaps) and "what" (semantic concepts) to "how" objects are spatially arranged relative to each other; proposes a taxonomy of spatial/semantic/scale context, reviews spatial XAI works, and calls for spatial-context benchmarks and measures.
Key results
- Five popular XAI methods (GradientSHAP, Integrated gradients, Occlusion, Saliency, LIME) produce vague, hard-to-interpret explanations for fine-tuned ResNet-50 and ViT-B on KITTI distance and dSprites orientation tasks.
- The paper finds only a few works on spatial XAI compared with many DL models that use spatial context.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Corrupt the order of input patches and pixels within patches | MLP-Mixer and ResNet · Images with corrupted spatial relationships | performance | Mixed |
Surprising findings
- Explanations for correct predictions look nearly identical to those that would be given for unrelated classes, so heatmaps cannot distinguish spatial-relation-driven decisions.
Limitations
- Evidence for XAI failures is qualitative and limited to two datasets and two models.
- Spatial XAI literature review is based on a limited survey of few works.
Experimental setup
Datasets: KITTI dSprites VTAB ImageNet
Benchmarks: KITTI (VTAB structured) dSprites (VTAB structured)
Metrics: qualitative attribution heatmap inspection
Baselines: GradientSHAP Integrated gradients Occlusion Saliency LIME
Builds on: LIME Grad-CAM Integrated gradients Concept Relevance Propagation CapsuleNet
Closest papers by meaning
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 2024 · similarity 0.81)
- Explainable Saliency: Articulating Reasoning with Contextual Prioritization (CVPR 2025 · similarity 0.81)
- Seeing the Unseen: Visual Common Sense for Semantic Placement (CVPR 2024 · similarity 0.81)
- Do Vision-Language Models Really Understand Visual Language? (ICML 2025 · similarity 0.81)
- ConTextual: Evaluating Context-Sensitive Text-Rich Visual Reasoning in Large Multimodal Models (ICML 2024 · similarity 0.80)
- Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models (NeurIPS 2026 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.