AI Paper Atlas

Research map / Detection and segmentation / Open-set and few-shot detection

OvarNet: Towards Open-Vocabulary Object Attribute Recognition

Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, Weidi Xie

CVPR 2023 · Detection and segmentation › Open-set and few-shot detection · approach: Open-vocabulary object detection · 46 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
Add parent-attribute words and learnable prompt vectors to the text prompt (ablation)CLIP-Attr · VAW box-givenAP_novel: 53.78 -> 57.39Improved
Finetune with COCO-base and VAW-base annotations (ablation)CLIP-Attr · VAW box-givenAP_novel: 46.15 -> 57.39Improved
Add image-caption training using COCO-Cap-sub (ablation)CLIP-Attr · VAW box-givenAP_all: 66.92 -> 69.03Improved
Add the caption, noun-phrase, and largest-box attribute supervision signals (ablation)CLIP-Attr Step-II · VAW box-givenAP_all: 66.92 -> 69.33Improved
Add top-K proposal attribute supervision to Step-II (ablation)CLIP-Attr Step-II · COCO box-givenAP_all: 63.80 -> 65.17Improved
Distill attribute prediction probabilities with KL loss (ablation)OvarNet · VAW box-givenAP_novel: 50.53 -> 56.43Improved
Use probability KL distillation instead of feature L2 distillation (ablation)OvarNet · COCO box-givenAP_novel: 33.17 -> 54.10Improved
Initialize the visual backbone from CLIP-Attr rather than ImageNet (ablation)OvarNet · VAW box-freeAP_novel: 50.27 -> 54.85Improved
Freeze rather than finetune the CLIP-Attr-initialized visual backbone (ablation)OvarNet · VAW box-freeAP_novel: 54.85 -> 55.47No clear effect
Replace CNN blocks with Transformer attentional pooling (ablation)OvarNet · VAW box-givenAP_novel: 53.35 -> 56.43Improved
Use 4 instead of 2 Transformer encoder layers for attentional pooling (ablation)OvarNet · VAW box-givenAP_novel: 55.02 -> 56.43Improved
Vary the number of learnable prompt vectors (ablation)CLIP-Attr · VAW box-givenAP_novel: 56.94 (3 prompts) -> 57.41 (60 prompts)No clear effect

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