AI Paper Atlas

Research map / Detection and segmentation / CNN features

Feature-Proxy Transformer for Few-Shot Segmentation

Jian-Wei Zhang, Yifan Sun, Yi Yang, Wei Chen

NeurIPS 2022 · Detection and segmentation › CNN features · approach: Context modeling for semantic segmentation · 26 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 pairwise loss that pulls foreground features together (ablation)FPTrans · PASCAL-5^i and COCO-20^i 1-shotmIoU: +1.1% PASCAL-5^i; +0.4% COCO-20^iImproved
Add query-support prompting and interaction (ablation)FPTrans with pairwise loss · PASCAL-5^i and COCO-20^i 1-shotmIoU: +1.0% PASCAL-5^i; +2.7% COCO-20^iImproved
Add multiple local background proxies (ablation)FPTrans with pairwise loss · PASCAL-5^i and COCO-20^i 1-shotmIoU: +1.1% PASCAL-5^i; +1.5% COCO-20^iImproved
Use all three components: pairwise loss, prompts, and multiple local proxies (ablation)FPTrans · PASCAL-5^i and COCO-20^i 1-shotmIoU: +2.9% PASCAL-5^i; 3.6% COCO-20^iImproved
Use learnable prompts without extracted prompts (ablation)FPTrans · PASCAL-5^i and COCO-20^i 1-shotmIoU: 64.0 -> 63.5; 40.3 -> 39.2Hurt
Add learnable vectors to extracted prompts (ablation)FPTrans · PASCAL-5^i and COCO-20^i 1-shotmIoU: 64.3 -> 64.7; 41.4 -> 42.0Improved
Include background-background pairs in the pairwise loss (ablation)FPTrans · PASCAL-5^i 1-shotmIoU: 64.7 -> 64.0 (50%); 64.7 -> 63.8 (100%)Hurt
Increase local background proxy count from 1 to 5 (ablation)FPTrans · PASCAL-5^i and COCO-20^i 1-shotmIoU: 63.9 -> 64.7; 41.5 -> 42.0Improved
Increase local background proxy count from 5 to 9 (ablation)FPTrans · PASCAL-5^i and COCO-20^i 1-shotmIoU: 64.7 -> 64.0; 42.0 -> 41.2Hurt
Replace CNN backbones with transformer backbones in decoder-based methods (ablation)PFENet and CyCTR · PASCAL-5^i 1-shotmIoU: PFENet: 60.8 -> 58.7 (ViT), 57.7 (DeiT); CyCTR: 64.2 -> 60.1 (ViT), 61.0 (DeiT)Hurt
Use FPTrans instead of HSNet under COCO-to-PASCAL domain shiftFPTrans · COCO-20^i→PASCAL-5^i 1-shot and 5-shotmIoU: +5.6%; +9.0%Improved

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