AI Paper Atlas

Research map / Detection and segmentation / Weakly supervised detection and segmentation

SegRCDB: Semantic Segmentation via Formula-Driven Supervised Learning

Risa Shinoda, Ryo Hayamizu, Kodai Nakashima, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka

ICCV 2023 · Detection and segmentation › Weakly supervised detection and segmentation · approach: Weakly supervised semantic segmentation · 12 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
pre-train with formula-driven SegRCDB instead of COCO-Stuff at same image countSwin-B + UPerNet · ADE-20kmIoU: 43.39 -> 43.85Improved
pre-train with SegRCDB instead of from scratchSwin-B + UPerNet · CityscapesmIoU: 54.65 -> 73.06Improved
use ImageNet backbone pre-training instead of SegRCDBSwin-B + UPerNet · ADE-20kmIoU: 43.85 -> 46.37Improved
use grayscale instead of random class colors (ablation)SegRCDB · ADE-20kmIoU: 28.49 -> 34.05Improved
increase occlusion range r from 100 to 400 (ablation)SegRCDB · ADE-20kmIoU: 30.62 -> 34.12Improved
increase categories from 64 to 255, then to 500 (ablation)SegRCDB · ADE-20kmIoU: 31.67 -> 34.05 (255) -> 31.58 (500)Mixed
increase number of images from 20k to 80k (ablation)SegRCDB · ADE-20kmIoU: 41.31 -> 42.25Improved
scale SegRCDB from 118k to 1M images (ablation)SegRCDB · ADE-20k vs CityscapesmIoU: ADE-20k 43.85 -> 44.46; Cityscapes 73.06 -> 72.10Mixed
vary line width (1 to 3 px) and polygon count (ablation)SegRCDB · ADE-20kmIoU: 34.05 vs 34.03 (line width)No clear effect
use Swin-B instead of ResNet-101 (ablation)SegRCDB · CityscapesmIoU: 66.74 -> 70.23Improved
shift annotation vertices by 100 pixels (ablation)SegRCDB · CityscapesmIoU: 70.23 -> 63.69Hurt
inflate polygons by 100 pixels (ablation)SegRCDB · CityscapesmIoU: 70.23 -> 67.30Hurt

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