Source-linked AI summary
Beyond Mean Foils: Auditing Worst-Foil Specificity in Frozen CLIP Region Explanations
Kaixin Liu, Zhipeng Ye, Feng Jiang, Zhenghao Wang, Qihang Wu
TL;DR
A region selected to explain a target in frozen CLIP can still contribute more to a competing class, leaving evidence about specificity incomplete when only overlap and mean-foil contrast are checked. The paper audits CCI regions against the strongest foil and tests alternative regions under a target-drop constraint. Many screened regions fail the worst-foil test, and only a small fraction of failures have a passing alternative within the stated tolerance.
Problem
A region can overlap a target and beat the average competing class yet contribute more to one competitor, so target specificity requires testing against the strongest foil.
Source: §1 · p. 1Method
Results
Only 0.16–0.98% of screened failures are repairable at ϵ = .02, despite passing alternatives for 6.25–7.84% of COCO and 27.40–31.15% of VOC failures without that constraint.
Source: §4 · p. 3Takeaways & Limitations
Worst-foil comparisons add information beyond overlap and mean contrast, while available regions and target-drop tolerance constrain repair opportunities.
Source: §4 · p. 3, §5 · p. 4Among repaired failures, mean bbox precision falls by 14.04–28.17 percentage points, and VOC estimates summarize only five and two repairs.
Source: §4 · p. 4
Abstract
from arXiv · showhide
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 $\epsilon = 0.02$ reduces these rates to 0.16-0.98%. Available regions and target-drop tolerance constrain repair; relaxing the tolerance increases repair opportunities.
1 INTRODUCTION
The introduction asks whether CCI-selected CLIP regions are truly target-specific: a region can overlap the target and beat the average foil while affecting one competing class more. The audit compares the strongest masked-response foil and considers whether another region can pass while preserving target response.
1 INTRODUCTION
- CCI groups patches into regions and measures how masking them changes image–text scores; CCI-top1 is the region whose removal most reduces the target score.The paper explains CLIP with Cluster-based Concept Importance (CCI) [2].Source: §1 · p. 1
- A region may contribute more to a competing class than to the target even when it overlaps the target and beats the average competing drop.The audit tests against the strongest competing class, or worst foil.Source: §1 · p. 1
- Unlike class-contrastive explanations that choose a competitor by its original-image score, this audit selects the class with the largest score drop when the region is masked.It also controls for competing objects and foil count, then examines all regions for repair constraints.Source: §1 · p. 1
2 AUDIT AND REPAIR OPPORTUNITY
The audit evaluates a fixed CCI-top1 region against the strongest competing-class drop using three evaluation prompts, then tests whether worst-foil reranking can find an alternative within a target-drop budget. It also uses an oracle and foil-removal controls to distinguish candidate availability from ranking and annotation effects.
2 AUDIT AND REPAIR OPPORTUNITY
- The response dR(c) is the original pre-softmax score minus the score after attention to region R is masked, with positive values indicating a reduction.The score includes CLIP’s learned logit scale, and masking retains the class-token attention column.Source: §2 · p. 1
- Worst-foil reranking selects a candidate whose target-drop loss is at most ϵ relative to CCI-top1, maximizing its selection-prompt margin against the largest foil drop.The screened baseline fixes images that pass both bbox precision of at least .5 and target drop greater than mean foil drop; failures have P < 0.Source: §2 · pp. 1–2
- The evaluation fixes the selected region before testing specificity, target response, and locality across three prompts, separating selection-stage responses from held-out evaluation.Worst-foil margin subtracts the largest foil drop from the target drop; a negative margin means a foil loses more score than the target.Source: §2 · p. 1, Figure 1 · p. 2
- The oracle counts images with a feasible candidate passing the held-out margin test, while the unrestricted version removes the budget and searches all K candidates.The budgeted oracle uses the feasible set Rϵ(i); the paper uses ϵ = .02 logit units and sweeps .01, .02, .05, .10, and .20.Source: §2 · p. 2
- The unrestricted oracle tests all K candidates and counts images with a feasible region whose held-out margin is nonnegative; feasible repairs and ranking misses are distinguished from images with no passing candidate.The four groups partition the screened failure set; NO = C2 + r, while the joint count j additionally requires overlap, mean-foil checks, and positive target contribution.Source: §2 · p. 2
- To isolate annotated-object effects, the audit removes annotated-present classes from each foil set and compares the resulting failure indicator with random subsets retaining the same number of foils.Negative ∆A favors annotation-based removal, and retained foil sets are nonempty.Source: §2 · p. 2
3 PROTOCOL AND REPRODUCIBILITY
The protocol evaluates frozen CLIP checkpoints on COCO and VOC image records, with targets and foils determined from annotations. It specifies the patch-region construction and attention intervention used for reproducibility.
3 PROTOCOL AND REPRODUCIBILITY
- The evaluation covers 27,708 COCO val2014 and 1,943 VOC2007 test images with CLIP B/16 and B/32, totaling 59,302 image–model records.Targets are the largest instances by annotation area on COCO or box area on VOC; images require another category and target area of 2–80%, and crowd annotations are excluded.Source: §3 · p. 2
- Each image has K = 8 patch regions formed by K-means over normalized projected patch features, and the intervention masks selected patch-key attention logits before softmax.K-means uses three starts, 50 iterations, and seed 1701 + i; intervention retains the class-token column, and boxes follow the 224-pixel resize and center crop.Source: §3 · p. 2
4 RESULTS AND ANALYSIS
Removing annotated competitors leaves most screened failures, and only a small share can be repaired with one of the eight candidate regions under the target-drop budget. Relaxing that budget increases repair opportunities, but does not uniformly improve target response or localization.
Competitor exclusion
- 96.63–98.24% of screened failures remain after removing annotated competitors; exclusion helps more than matched random removal on COCO, while VOC confidence intervals include zero.The results attribute only a small part of the observed failure to annotated competing objects.Source: §4 · p. 3, Table 1 · p. 3
Positive target contribution and scene density
- After exclusion, failure remains 61.38–62.62% on COCO and 39.26–40.05% on VOC even when target contribution is positive.COCO images with exactly two annotated categories still have failure rates of 63.39% and 64.95% for B/16 and B/32, respectively.Source: §4 · p. 3
Repair capacity
- Only 6.25–7.84% of COCO failures and 27.40–31.15% of VOC failures have a passing alternative; with ϵ = .02, just 0.16–0.98% are repairable.The main losses occur before the final ranking decision; WF misses at most one feasible repair per setting.Source: Figure 2 · p. 3, §4 · p. 3, Table 2 · p. 3
Budget sensitivity
- At ϵ = .20, oracle repair capacity rises to 1.57–2.05% of COCO failures and 6.65–7.14% of VOC failures.Multiple feasible candidates become available on about 35% of eligible images; at .20, raw target drops decrease in all four settings, while overlap decreases on COCO and increases on VOC.Source: Table 3 · p. 3, §4 · p. 3
Normalization sensitivity
- Normalization changes failure rates by at most 0.083 percentage points, and the worst foil remains unchanged on 99.42–99.84% of images.The comparison fixes earlier-run regions and images rather than pooling those records with rerun results.Source: Table 4 · p. 4, §4 · p. 4
Repair trade-offs
- Among repaired failures, mean bbox precision falls by 14.04–28.17 percentage points; its interval spans zero for VOC B/32, and VOC estimates use only five and two repairs.Mean target-score changes are positive on COCO and negative on VOC, but intervals span zero for both B/16 settings.Source: Table 5 · p. 4, §4 · p. 4
Retaining the original checks
- A repaired margin does not always retain overlap, mean-foil contrast, and positive target contribution: on COCO B/32, 30 repairs include only 25 passing all checks.The oracle finds 12, 26, 3, and 2 feasible repairs passing all checks, in table order.Source: §4 · p. 4, Table 5 · p. 4
Failure thresholds and uncertainty
- At δ = .10, 59.49–60.22% of COCO screen-passers and 34.39–36.23% of VOC screen-passers still fail; rates decline further at .50.The main results use a recorded rerun of the earlier seed rule, without independent base-seed repetitions.Source: Table 6 · p. 4, §4 · p. 4
5 CONCLUSION
Worst-foil comparisons reveal failures that overlap and mean contrast alone miss in frozen CLIP region explanations. Repair is constrained by the available regions and target-drop tolerance, and relaxing that tolerance changes target response and overlap.
5 CONCLUSION
- Worst-foil comparisons reveal frequent CCI-top1 failures despite target overlap and mean contrast; many images lack a passing alternative, while most remaining alternatives exceed the target-drop tolerance.Relaxing the tolerance allows more repairs but changes target response and overlap.Source: §5 · p. 4
COMPLIANCE WITH ETHICAL STANDARDS
The study uses publicly released COCO and PASCAL VOC images and annotations, with no recruited participants or new human or animal experiments.
COMPLIANCE WITH ETHICAL STANDARDS
- The work analyzes publicly released COCO and PASCAL VOC images and annotations, with no recruited participants or new human or animal experiments.Source: COMPLIANCE WITH ETHICAL STANDARDS · p. 5