AI Paper Atlas

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

Prototypical Kernel Learning and Open-set Foreground Perception for Generalized Few-shot Semantic Segmentation

Kai Huang, Feigege Wang, Ye Xi, Yutao Gao

ICCV 2023 · Detection and segmentation › Open-set and few-shot detection · approach: Few-shot segmentation · 18 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 prototypical kernel learning (PKL) to the baseline (ablation)GFSS kernel baseline · PASCAL-5i 1-shothIoU: 23.45 -> 34.17Improved
add foreground contextual perception (FCP) to the baseline (ablation)GFSS kernel baseline · PASCAL-5i 1-shothIoU: 23.45 -> 26.57Improved
add conditional bias based inference (CBBI) on top of PKL+FCP (ablation)Ours · PASCAL-5i 1-shothIoU: 34.98 -> 37.83Improved
add CBBI on top of PKL+FCP (ablation)Ours · PASCAL-5i 5-shothIoU: 43.02 -> 45.42Improved
use adaptive learning rate (cosine similarity with ReLU) instead of fixed rate in kernel update (ablation)PKL · PASCAL-5i 1-shothIoU: 31.36 -> 37.83Improved
use prototypical kernel update instead of concat plus linear projection or adaptive pooling fusion (ablation)PKL · PASCAL-5i 1-shothIoU: 32.11 -> 37.83 (vs concat+LP)Improved
use max pooling over pixels and average pooling over images in FCP hybrid pooling (ablation)FCP · PASCAL-5i 1-shothIoU: 33.34 -> 37.83 (pixel avg pooling variant)Improved
use FCP instead of salient detectors PAGENet or MINet for foreground (ablation)Ours · PASCAL-5i 1-shothIoU: 36.28 -> 37.83 (vs PAGENet); 36.53 (MINet)Improved
use GradCAM as foreground module instead of FCP (ablation)Ours · PASCAL-5i 1-shothIoU: 37.83 -> 34.74Hurt
use unconditional or learnable-convolution fusion instead of conditional bias inference (ablation)Ours · PASCAL-5i 1-shothIoU: uCBBI 31.70, cCBBI 29.29 vs 37.83; both below no CBBI (34.98)Hurt
increase the number of COCO classes (15 to 60 base classes) (ablation)Ours versus CAPL · COCO subset, novel mIoU gain over CAPLmIoU novel improvement: +7.76 -> +3.99Hurt
replace CAPL with the proposed method in GFSS on PASCAL-5iGFSS PASCAL-5i · PASCAL-5i 5-shothIoU: 31.96 -> 45.42Improved

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