Sample adaptive token routing improves 3D model test time accuracy
SATOR: Backpropagation-Free 3D Test-Time Adaptation via Sample-Adaptive Token Routing
Computer Vision and Pattern Recognition
Summary
3D models trained on one set of data can perform worse when tested on new, different data. The authors created a method called SATOR that helps these models adapt without needing extra training steps during testing. SATOR smartly decides which parts of the 3D input to keep or ignore based on each sample’s unique data. This approach makes the model more accurate when handling corrupted or changed 3D shapes.
What this means in practice
- •For 3d computer vision engineers: Improve 3D object recognition accuracy in robots and autonomous systems facing varied or corrupted sensor data by adapting models at test time without retraining.
- •For augmented reality developers: Enhance real-time adaptation of 3D scene understanding models to unseen environmental changes, improving user experience under diverse conditions.
Authors
Inseok Kong, Geunyoung Jung, Jiyoung Jung
Abstract
Test-time adaptation (TTA) mitigates the performance degradation of source-pretrained models on unlabeled test data under distribution shifts. Recent backpropagation-free token-purging methods address 3D point cloud distribution shifts by filtering unreliable tokens. However, their reliance on predefined purge-size candidates requires multiple forward passes per sample and increases inference cost. We propose SATOR (Sample-Adaptive TOken Routing), a source-free and backpropagation-free 3D TTA framework that eliminates this search by deriving both the partitioning threshold and routing decision directly from each sample. SATOR computes token abnormality scores from complementary attention and geometric relation graphs. It then derives a sample-specific partition into retained and pruned tokens via Otsu-based separability, without evaluating candidate sizes. Keep-ratio-calibrated routing then selects the pruned prediction only when its entropy, penalized by the keep ratio, is lower than that of the original prediction. Experiments on ModelNet40-C, ScanObjectNN-C, and ShapeNet-C demonstrate that SATOR achieves state-of-the-art accuracy on three 3D corruption benchmarks.