AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

Chatting Makes Perfect: Chat-based Image Retrieval

Matan Levy, Rami Ben-Ari, Nir Darshan, Dani Lischinski

NeurIPS 2023 · Vision-language and multimodal › Human and embodied modeling · approach: Neural image captioning · 7 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train retriever on dialogues instead of using off-the-shelf BLIP on captionChatIR · COCOretrieval accuracy (single-hop): 83\% to 87\%Improved
add dialogue rounds to the query for fine-tuned BLIP baselineBLIP / CLIP · Flickr30K and COCOretrieval accuracyImproved
ChatIR dialogue-trained model vs BLIP fine-tuned single-hop at 0 roundsChatIR · VisDial valretrieval performance: 63.66\% vs. 63.61\%No clear effect
add 2 dialogue rounds with ChatGPT questionerChatIR · VisDial val, 50K imagesHit rate@10: 63% to 73.5%Improved
ChatGPT questioner with full chat history instead of Human questionsChatIR · VisDial valHit rate@10: ~2%Improved
ChatGPT questioner without access to answers (Unanswered) vs Human (ablation)ChatIR · VisDial valHit rate@10: ~0.5%No clear effect
use RL-based questioners or FLAN models instead of ChatGPT (ablation)ChatIR · VisDial valHit rate@10 / repeated questions: 1.85-3.44Hurt
replace BLIP2 answerer with Human answerer (ablation)ChatIR with ChatGPT questioner · VisDial val subsetHit rate@10: 81%Improved
mask image captions during retriever training (ablation)Image Retriever F · VisDialretrieval performance: 2-3%Improved
mask questions, answers, Q&A rounds or tokens instead of captions (ablation)Image Retriever F · VisDialretrieval performanceNo clear effect
augment training with a second synthetic dialogue per image (ablation)Image Retriever F · VisDialHits@10 and ATRMixed
add dialogues on 73K new COCO images (68% more images) (ablation)Image Retriever F · VisDialATR: 4\% drop in ATRHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy