Zero-Shot Day-Night Domain Adaptation with a Physics Prior
This addresses the challenge of adapting models from day to night domains without costly test data, which is incremental as it builds on existing domain adaptation methods by incorporating physics priors.
The paper tackles the problem of zero-shot day-night domain adaptation by using a physics-based reflection model as a visual inductive prior, eliminating the need for test data imagery. It demonstrates improved performance on synthetic and natural datasets for tasks like classification, segmentation, and place recognition, with reduced distribution shift in feature maps.
We explore the zero-shot setting for day-night domain adaptation. The traditional domain adaptation setting is to train on one domain and adapt to the target domain by exploiting unlabeled data samples from the test set. As gathering relevant test data is expensive and sometimes even impossible, we remove any reliance on test data imagery and instead exploit a visual inductive prior derived from physics-based reflection models for domain adaptation. We cast a number of color invariant edge detectors as trainable layers in a convolutional neural network and evaluate their robustness to illumination changes. We show that the color invariant layer reduces the day-night distribution shift in feature map activations throughout the network. We demonstrate improved performance for zero-shot day to night domain adaptation on both synthetic as well as natural datasets in various tasks, including classification, segmentation and place recognition.