Test-time adaptation (TTA) helps source-pretrained models cope with unlabeled test data that differs from the training distribution. Conventional TTA methods rely on backpropagation to update model parameters, which is computationally expensive and often impractical for real-time or resource-constrained settings.
The new paper introduces SATOR, a backpropagation-free TTA approach designed for 3D point cloud data. It builds on recent token-purging methods that adapt models by selectively discarding tokens, but adds a sample-adaptive routing mechanism to decide which tokens to keep for each individual test sample.
The abstract is truncated, so details on the routing algorithm and experimental results are not fully available in the source. Still, the contribution is clear: a computationally efficient alternative to gradient-based TTA for 3D point clouds, potentially broadening the applicability of adaptation methods in real-world scenarios.