11 DAILY CVPR Wednesday The team’s challenge was to surpass state-of-the-art performance with their new method. The paper demonstrates that their approach can recover classification performance significantly compared to existing TTA baselines. Noting that this is still a relatively new field, David and Gustavo are eager to share their work with the broader research community at CVPR today. Their poster session is an opportunity to introduce NC-TTT and foster discussions about the future of TTA and TTT in general. “I’d invite any deep learning researcher to think about how a model they train can fail at test time,” Gustavo urges. “This is just one alternative we can provide to prepare the model before that happens. The most exciting part is coming up with new auxiliary tasks that work better and better!” Looking ahead, David sees their work as the beginning of a broader exploration of TTA. “It’s not the end of something,” he surmises. “We’ve seen a lot of TTT and TTA articles. Many fields are progressing in TTA, like working on Vision-Language Models (VLMs) or segmentation. I think it’s something that will be everywhere in the future – this is just the beginning.” To learn more about David and Gustavo’s work, visit Poster Session 2 & Exhibit Hall (Arch 4A-E) from 17:15 to 18:45 [Poster 123]. NC-TTT
RkJQdWJsaXNoZXIy NTc3NzU=