Frozen clip region explanations often misattribute objects to wrong classes
Beyond Mean Foils: Auditing Worst-Foil Specificity in Frozen CLIP Region Explanations
Computer Vision and Pattern Recognition
Summary
Sometimes parts of an image that belong to one object are mistakenly shown as more related to another object in visual AI explanations. The authors studied these mix-ups in a popular image-text model called frozen CLIP and found that many explanations fail to clearly identify the right object. They also tested ways to fix these errors, but improvements were limited because of strict requirements for explanation accuracy. Their work helps reveal the limits of current AI explanations in distinguishing closely related objects in images.
What this means in practice
- •For computer vision engineers: Improve interpretability tools by identifying and reducing misattributed regions in frozen CLIP image explanations for clearer object understanding.
- •For machine learning auditors: Enhance auditing processes by testing and repairing regional explanations to better detect when models confuse similar classes in image tasks.
Authors
Kaixin Liu, Zhipeng Ye, Feng Jiang, Zhenghao Wang, Qihang Wu
Abstract
A region can overlap a target object yet contribute more to another class. We test regions selected by Cluster-based Concept Importance (CCI) in frozen CLIP. Across COCO and VOC with two checkpoints, 41.08-64.78% of regions that pass overlap and mean-contrast checks fail against the strongest competing class. Removing competitors annotated in the image leaves 39.69-63.64% failing. We then test all eight candidate regions per image. An alternative passes the test for 6.25-7.84% of failures on COCO and 27.40-31.15% on VOC. Requiring it to preserve the original target-score drop within $ε= 0.02$ reduces these rates to 0.16-0.98%. Available regions and target-drop tolerance constrain repair; relaxing the tolerance increases repair opportunities.