IEEE/CVF WACV 2027

SATOR: Backpropagation-Free 3D Test-Time Adaptation
via Sample-Adaptive Token Routing

Inseok Kong1, Geunyoung Jung2, Jiyoung Jung2,*
1 Department of GeoInformatics, University of Seoul  •  2 Department of Artificial Intelligence, University of Seoul
* Corresponding author
SATOR Framework Architecture Overview

Overview of the proposed SATOR framework. SATOR introduces a source-free and backpropagation-free 3D TTA architecture that computes token abnormality scores from complementary attention and geometric relation graphs, adaptively derives sample-specific token partitions via Otsu-based separability, and routes predictions using keep-ratio-calibrated routing without evaluating candidate purge sizes.

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. Extensive experiments on ModelNet40-C, ScanObjectNN-C, and ShapeNet-C demonstrate that SATOR achieves state-of-the-art accuracy on three 3D corruption benchmarks.

Key Highlights & Contributions

BP-Free
Zero Backward Passes Required
Search-Free
Zero Candidate Purge-Size Search
SOTA
ModelNet40-C, ScanObjectNN-C, ShapeNet-C
Backpropagation-Free 3D Adaptation

Conventional 3D TTA approaches rely on gradient-based backpropagation during inference, incurring excessive computational cost and memory footprints. SATOR operates completely without backward passes or source data, making it highly suitable for resource-constrained edge systems.

Complementary Attention & Geometric Relation Graphs

Instead of relying on stored global statistics or self-attention alone, SATOR synergizes internal self-attention maps with spatial geometric relation graphs ($K_g$-NN affinity) to quantify how corruptions disrupt local neighborhood structures and reliably score abnormal tokens.

Otsu-Based Search-Free Token Partitioning

Prior token-purging methods evaluate multiple predefined purge-size candidates through iterative forward passes. SATOR completely eliminates candidate search by deriving an optimal sample-specific threshold directly from the score distribution via Otsu-based separability.

Keep-Ratio-Calibrated Dual-Branch Routing

SATOR compares predictions from the original sequence and the pruned sequence. To prevent low entropy from being favored solely due to aggressive token reduction, it calibrates prediction entropy by the keep ratio ($\rho = |\mathcal{K}|/M$), selecting the pruned branch only when $\frac{H(z_{\mathrm{prune}})}{\rho} < H(z_{\mathrm{orig}})$.

Main Results

Mean accuracy under 15 corruption types

Method Venue ModelNet40-C ShapeNet-C ScanObjectNN-C
Source-only–54.1661.3536.96
TENTICLR'2163.5354.9843.45
SARICLR'2359.6460.1748.78
BFTT3DCVPR'2457.1660.7333.00
PGICCV'2568.6864.4046.78
LN-TTAWACV'2669.7164.3548.50
SATOR (Ours)–70.26 (+16.10)64.92 (+3.57)49.43 (+12.47)

Top-1 accuracy (%) averaged over 15 corruption types. Values in parentheses are gains over source-only. Best and second best are marked. Full per-corruption results and more baselines are in the paper.

Runtime and memory (ScanObjectNN-C)

Method Acc. (%) FPS Mem. (GB)
BFTT3D33.0055.98.77
PG46.78174.11.38
LN-TTA48.50790.01.38
SATOR49.43483.21.38

Skipping the purge-size search makes SATOR about 2.8× faster than PG, with higher accuracy and the same memory use.

Noise token localization

Score AUROC AP
Attention only0.7510.765
Geometry only0.7940.683
Adaptive fusion0.8390.783

On background corruption (ScanObjectNN-C), fusing the attention and geometry scores locates corrupted tokens best.

BibTeX

@inproceedings{kong2027sator,
  title={SATOR: Backpropagation-Free 3D Test-Time Adaptation via Sample-Adaptive Token Routing},
  author={Kong, Inseok and Jung, Geunyoung and Jung, Jiyoung},
  booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2027}
}