AI Paper Atlas

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

Rounded or Streamlined Head? Bridging Concept Bottleneck Models and Attribute-Described Object Parts

Yang Liu, Jiajin Zhang, Yaojun Hu, Bingguang Hao, Xin Cao, Yingda Xia, Danyang Tu, Shi Gu, Ling Zhang

CVPR 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Concept bottleneck models

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 cost aggregation module (ablation)OA-CBM · CUB-200 (patch token)top-1 accuracy: 33.8 -> 63.2Improved
add hierarchical clustering on top of cost aggregation (ablation)OA-CBM · PartAttrImageNet (Pred-All)h-IoU: 12.9 -> 35.7Improved
add hierarchical clustering on top of cost aggregation (ablation)OA-CBM · PartAttrCUB (Pred-All)h-IoU: 6.6 -> 20.8Improved
add hierarchical clustering on top of cost aggregation (ablation)OA-CBM · CUB-200 and CIFAR-100 classificationaccuracy: 63.2 -> 63.3 (CUB); 96.2 -> 96.1 (CIFAR-100)No clear effect
add hierarchical clustering on top of cost aggregation (ablation)OA-CBM · FOOD-101 and FLOWER classificationaccuracy: 87.7 -> 87.3 (FOOD-101); 92.8 -> 91.8 (FLOWER)Hurt
add hierarchical clustering without cost aggregation (ablation)OA-CBM · CUB-200 (patch token)top-1 accuracy: 33.8 -> 0.01Hurt
add object-aware clustering for Pred-All grounding (ablation)OA-CBM · Oracle-Obj/Oracle-All groundingh-IoU: identical with and without (e.g. 51.4 vs 51.4)No clear effect
replace CLIP text encoder with Llama3 and add part-attribute conceptsOA-CBM · ImageNet (CLS token)top-1 accuracy: 78.7 -> 83.1 vs DOT-CBMImproved
use patch-token concept scores instead of CLS-token scoresCBMs · classificationaccuracyHurt
use patch-guided hybrid scoring (CLS x sigmoid(patch))OA-CBM · CUB-200top-1 accuracy: 63.3 -> 80.6 vs patch-tokenImproved
evaluate grounding with attributes removed (part-only)CLIP and DINO baselines · PartAttr datasetsmIoUHurt

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