10 DAILY CVPR Wednesday changing from black and white to color can represent a domain shift for a model trained in black and white. “For an easy example, say you train your model with images of cats and dogs, and you want to classify it,” David illustrates. “You’ve only taken pictures at training during the spring, so then at test time, you’ll show some images that you took in the winter, and your model’s performance will decrease because it never saw images in winter. The goal of our architecture is to adapt to the winter images because it already has this knowledge at spring.” Gustavo tells us that TTA should be an essential component of any model deployed in the real world because models often encounter images that differ from their training data. “There will always be images that might be slightly different, or the camera that’s used to get the new images will change,” he points out. “At some point, the model’s performance will decrease, and the only way to ensure the model keeps going well is to have an adaptation mechanism that can work in real time or, in this case, at test time.” Highlight Presentation
RkJQdWJsaXNoZXIy NTc3NzU=