Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval

Researchers introduce SaMer, a framework that compresses image tokens for vision-language retrieval while preserving object-level evidence. SaMer reduces storage costs by 16.09x and improves performance on benchmarks like Flickr30K and MSCOCO. This approach maintains query-selectable object evidence, outperforming existing compression methods.

1sec.ai