The paper "Test-Time Adaptation for Depth Completion" addresses the challenge of performance degradation when transferring models trained on source datasets to target datasets due to domain gaps. The authors propose an online test-time adaptation (TTA) method for depth completion, which aims to bridge the performance gap in a single pass without requiring access to the source data. The method leverages the observation that the sparse depth modality exhibits a smaller covariate shift compared to the image modality. By designing an embedding module trained in the source domain, the method preserves a mapping from sparse depth features to both image and sparse depth features. During test time, sparse depth features are projected using this mapping as a proxy for source domain features and are used to train auxiliary parameters (adaptation layer) to align image and sparse depth features from the target test domain to the source domain. The method is evaluated on indoor and outdoor scenarios and shows an average improvement of 21.1% over baselines. The authors also release code, models, and dataset benchmarking setup to facilitate further research.The paper "Test-Time Adaptation for Depth Completion" addresses the challenge of performance degradation when transferring models trained on source datasets to target datasets due to domain gaps. The authors propose an online test-time adaptation (TTA) method for depth completion, which aims to bridge the performance gap in a single pass without requiring access to the source data. The method leverages the observation that the sparse depth modality exhibits a smaller covariate shift compared to the image modality. By designing an embedding module trained in the source domain, the method preserves a mapping from sparse depth features to both image and sparse depth features. During test time, sparse depth features are projected using this mapping as a proxy for source domain features and are used to train auxiliary parameters (adaptation layer) to align image and sparse depth features from the target test domain to the source domain. The method is evaluated on indoor and outdoor scenarios and shows an average improvement of 21.1% over baselines. The authors also release code, models, and dataset benchmarking setup to facilitate further research.