AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Weakly Supervised Posture Mining for Fine-Grained Classification

Zhenchao Tang, Hualin Yang, Calvin Yu-Chian Chen

CVPR 2023 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 40 citations (OpenAlex)

Read the paper

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
replace direct node-feature concatenation with message passing (ablation)PMRC (ResNet50, 6 regions) · CUB-200-2011Top-1 accuracy (%): +2.7 (91.8 vs 89.1)Improved
replace CE with RCE loss for the whole-graph classification (ablation)PMRC (ResNet50, 6 regions) · CUB-200-2011Top-1 accuracy (%): +1.2 (91.8 vs 90.6)Improved
replace CE with RCE loss (ablation)PMRC (SwinTransformer) · CUB-200-2011Top-1 accuracy (%): +0.8 (94.3 vs 93.5)Improved
replace CE with RCE loss when training the plain backbone (ablation)ResNet50 backbone · CUB-200-2011Top-1 accuracy (%): +1.2 (85.7 vs 84.5)Improved
apply RCE instead of CE to a range of existing fine-grained modelsResNet50, DenseNet161, Xception, Inception-V3, MobileNetV2, B-CNN, NTS-Net and others · CUB, Cars, Aircraft, DogsTop-1 accuracy increment vs CE: mostly +0.4 to +2.1; DATL -0.28 on CUB, -0.19 on Aircraft, TResNet-L-V2 -0.07 on CUBMixed
add RCE loss and posture-mining graph (weakly supervised, image labels only)PMRC (SwinTransformer) vs SwinTransformer backbone · CUB-200-2011Top-1 accuracy (%): +4.0 (94.3 vs 90.3)Improved
add PMRC graph/posture framework on ResNet50 backboneResNet50 · CUB-200-2011Top-1 accuracy (%): +7.3 (91.8 vs 84.5)Improved
varying number of discriminative regions (3 to 8) (ablation)PMRC (ResNet50, message passing, RCE) · CUB-200-2011Top-1 accuracy (%): peak at 6 regions (91.8); 3 regions 90.9, 8 regions 91.4Mixed

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