AI Paper Atlas

Research map / Detection and segmentation / CNN features

Exploring Cross-Image Pixel Contrast for Semantic Segmentation

Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, Luc Van Gool

ICCV 2021 · Detection and segmentation › CNN features · approach: Context modeling for semantic segmentation · 618 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 inter-image (cross-image) pixel contrast instead of no contrast (ablation)HRNetV2-W48 segmentation · Cityscapes valmIoU: 78.1 -> 81.0Improved
restrict contrast to pixels within the same image (intra-image) rather than across images (ablation)HRNetV2-W48 segmentation · Cityscapes valmIoU: 81.0 -> 78.9Hurt
add a memory bank to mini-batch-only contrast (ablation)pixel contrast with HRNetV2-W48 · Cityscapes valmIoU: 79.8 -> 80.5 (pixel memory), 79.8 -> 80.2 (region memory)Improved
add region memory to pixel memory (ablation)pixel contrast with HRNetV2-W48 · Cityscapes valmIoU: 80.5 -> 81.0Improved
use semi-hard rather than random positive/negative sampling (ablation)pixel contrast with random anchors · Cityscapes valmIoU: 79.3 -> 80.1Improved
use hardest rather than random positive/negative sampling (ablation)pixel contrast with random anchors · Cityscapes valmIoU: 79.3 -> 79.4No clear effect
use semi-hard rather than hardest sampling with hard anchors (ablation)pixel contrast with seg.-aware hard anchors · Cityscapes valmIoU: 80.5 -> 81.0Improved
add segmentation-aware hard anchor sampling to semi-hard sampling (ablation)pixel contrast · Cityscapes valmIoU: 80.1 -> 81.0Improved
add contrastive loss to cross-entropy, versus adding Lovasz loss (ablation)HRNetV2-W48 segmentation · Cityscapes valmIoU: 80.3 -> 81.0Improved
add contrastive loss on top of AAF loss (ablation)HRNetV2-W48 + CE + AAF · Cityscapes valmIoU: 81.0 -> 81.0No clear effect
add contrastive loss on top of RMI loss (ablation)HRNetV2-W48 + CE + RMI · Cityscapes valmIoU: 79.8 -> 81.3Improved
add contrastive loss on top of Lovasz loss (ablation)HRNetV2-W48 + CE + Lovasz · Cityscapes valmIoU: 80.3 -> 81.5Improved

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