サインイン
インサイト - Vision-Language
ALOHa: A Novel Metric for Detecting Object Hallucinations in Image Captions
ALOHa leverages large language models to reliably and localizeably detect object hallucinations in image captions, outperforming prior methods.
Efficient Language-Only Training for Zero-Shot Composed Image Retrieval
A novel language-only training framework, LinCIR, that efficiently learns a projection module to enable zero-shot composed image retrieval without relying on expensive image-text-image triplet datasets.
Semantics-enhanced Cross-modal Masked Image Modeling for Vision-Language Pre-training
The author proposes a SemMIM framework to enhance cross-modal semantic alignment by injecting high-level semantics into local patch encodings and involving text deeply in the MIM process.
The All-Seeing Project V2: Enhancing Relation Comprehension in Vision-Language Models
The author introduces the All-Seeing Project V2 to improve relation comprehension in vision-language models through a novel task called Relation Conversation (ReC).
