This repository contains the official Pytorch implementation of under review paper: "ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains"
Abstract
This paper introduces ReservoirTTA, a novel plug–in framework designed for prolonged test–time adaptation (TTA) in scenarios where the test domain continuously shifts over time, including cases where domains recur or evolve gradually. At its core, ReservoirTTA maintains a reservoir of domain-specialized models—an adaptive test-time model ensemble—that both detects new domains via online clustering over style features of incoming samples and routes each sample to the appropriate specialized model, and thereby enables domain-specific adaptation. This multi-model strategy overcomes key limitations of single model adaptation, such as catastrophic forgetting, inter-domain interference, and error accumulation, ensuring robust and stable performance on sustained non-stationary test distributions. Our theoretical analysis reveals key components that bound parameter variance and prevent model collapse, while our plug–in TTA module mitigates catastrophic forgetting of previously encountered domains. Extensive experiments on the classification corruption benchmarks, including ImageNet-C and CIFAR-10/100-C, as well as the Cityscapes→ACDC semantic segmentation task, covering recurring and continuously evolving domain shifts, demonstrate that ReservoirTTA significantly improves adaptation accuracy and maintains stable performance across prolonged, recurring shifts, outperforming state-of-the-art methods.
Overview
image
Recurring test-time adaptation scenarios. Left: Visual domains can recur over time; ETA, lacking regularization, steadily degrades under these repeated shifts. Right: A zoom-in on the snow corruption across 20 recurrences shows that EATA remains overall stable but still exhibits error spikes on returning to the same corruption across recurrences. ReservoirTTA detects returning domains and reuses specialized models to preserve learned knowledge, delivering improved robustness and faster (re-)adaptation over successive recurrences.
image
Overview of ReservoirTTA. ReservoirTTA operates in four stages: (1) Style Characterization and Domain Identification extracts early convolutional features and assigns incoming test batches to a style cluster via an online clustering mechanism; (2) Model Reservoir Initialization adds a new model for a detected domain, initializing it with parameters that maximize prediction mutual information; (3) Model Reservoir Adaptation selectively adapts the most relevant model using TTA methods; and (4) Model Prediction is then obtained via the ensemble’s parameters.
Prerequisites
To use the repository please use the following conda environment
@article{vray2025reservoirtta,
title={ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains},
author={Vray, Guillaume and Tomar, Devavrat and Gao, Xufeng and Thiran, Jean-Philippe and Shelhamer, Evan and Bozorgtabar, Behzad},
journal={arXiv preprint arXiv:2505.14511},
year={2025}
}
