FAST-GOAL: Fast and Efficient Global-local Object Alignment Learning

In three linesFAST-GOAL enhances CLIP to handle lengthy text descriptions through global-local semantic alignment. The method combines efficient local region extraction (FLISM) and token similarity-based learning (TSL). A new GLIT100k dataset with global image-caption pairs and derived local pairs validates the approach on DOCCI, DCI, MSCOCO, Flickr30k.

Summary generated by Claude — human-verified