Physical Knot Classification Beyond Accuracy: A Benchmark and Diagnostic Study

arXiv:2603.23286v3 Announce Type: replace Abstract: Physical knot classification is a fine-grained task in which the intended cue is rope crossing structure, but high accuracy may still come from appearance shortcuts. Using a tightness-stratified benchmark built from the public 10Knots dataset (1,440 images, 10 classes), we train on loose knots and test on tightly dressed knots to evaluate whether structure-guided training yields topology-specific gains. Topological distance predicts residual confusion for several backbones, but a random-distance control shows that topology-aware centroid alignment does not provide reliable topology-specific improvement. Auxiliary crossing-number prediction is more stable than centroid alignment, but neither method removes strong appearance reliance. In causal probes, background changes alone flip 17-32% of predictions, and phone-photo accuracy drops by 58-69 percentage points. These results show that high closed-set accuracy does not imply topology-dominant recognition, and that appearance bias remains the main obstacle to deployment.

Related

• ASBench: Image Anomalies Synthesis Benchmark for Anomaly Detection
• Action Without Interaction: Probing the Physical Foundations of Video LMMs via Contact-Release Detection
• Mitigating Spurious Background Bias in Multimedia Recognition with Disentangled Concept Bottlenecks
• AnomalyVFM -- Transforming Vision Foundation Models into Zero-Shot Anomaly Detectors

Source: arXiv cs.CV | 2026-04-10