Get Started
Home
Topics
Search
Library
Research questionHow can post-hoc vision saliency remain faithful when image orientation changes?A model may preserve its prediction after an image rotation while Grad-CAM shifts toward different regions. In histopathology and aerial imagery, this drift makes it difficult to distinguish evidence from artifacts in an explanation.
AI
Computer Vision
Image & Video Processing
Machine Learning
Research Paper
Latest papersRecent research connected to this question, newest first.Signal or Noise? Auditing Rotation-Induced Saliency Drift in Medical and Aerial ImagingThe work studies Grad-CAM-style explanations for rotated inputs across image classifiers, with experiments on ImageNet-1K, PatchCamelyon, and RESISC45, including ResNet-50, VGG-16, ViT-B/16, and zero-shot CLIP. It analyzes equivariance within the saliency operator and evaluates a training-free multi-view alignment-and-averaging wrapper; the evidence is empirical and concerns rotation-induced explanation drift.research paper · Sep 2, 2026
Related questions
How should self-supervised visual learning combine objectives to prevent collapse while preserving semantic and spatial information?How can visual attention distinguish class-relevant image features from closely associated distractors?How should landmark-derived features preserve task-relevant geometry under nuisance variation in structured image classification?How can vision systems discover unknown fine-grained classes without capture-device bias distorting cross-device evaluation?