Reproduce “Bridging CLIP with DINO: Cross-Modal Information Maximization for Online Test-Time Adaptation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Taehoon Kim, Wonjun Hwang, Hyunsouk Cho
CDIM adapts CLIP online to unlabeled corrupted images without source data. It combines CLIP text prototypes projected into frozen DINO features with a batch-diversity objective and prediction-frequency correction, while updating only CLIP’s normalization parameters. On ImageNet-C at severity 5, it achieves 36.9% mean top-1 accuracy, versus 32.1% for SubTTA; this result is limited to the tested benchmark and protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.