AI Paper Atlas

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

TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation

Zhaoyuan Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li, Rong Jin

ECCV 2022 · Detection and segmentation › Open-set and few-shot detection · approach: Unsupervised object and part discovery · 42 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 teacher-student bootstrapping iterations after training the student once (ablation)TransFGU · COCO-Stuff-171mIoU: 8.66 -> 11.93Improved
train a student on the initial pseudo labels (ablation)TransFGU · COCO-Stuff-171mIoU: 5.05 -> 8.66Improved
use ViT with Segmenter decoder instead of R50+FPN encoder (ablation)TransFGU · COCO-Stuff-171mIoU: 9.21 -> 11.93Improved
use Segmenter decoder instead of conv classifier on ViT (ablation)TransFGU · COCO-Stuff-171mIoU: 7.96 -> 11.93Improved
add peer loss to CE (ablation)TransFGU · COCO-Stuff-171mIoU: 10.49 -> 10.54Improved
add uncertainty loss to peer loss (ablation)TransFGU · COCO-Stuff-171mIoU: 10.54 -> 11.24Improved
add diversity loss to peer loss (ablation)TransFGU · COCO-Stuff-171mIoU: 10.54 -> 10.96Improved
combine peer, uncertainty and diversity losses (ablation)TransFGU · COCO-Stuff-171mIoU: 11.24 -> 11.93Improved
use DINO ViT-S/8 instead of ViT-S/16 as prior (ablation)TransFGU · COCO-Stuff-171mIoU: 9.43 -> 11.93Improved
use DINO instead of MoCoV3 as the prior (ablation)TransFGU · COCO-Stuff-171mIoU: 2.32 -> 9.43Improved
separate foreground and background patches before clustering (ablation)TransFGU · COCO-Stuff-171mIoU: 9.52 -> 11.93Improved
use ground-truth crops instead of sliding windows (ablation)TransFGU · COCO-Stuff-171mIoU: 11.93 -> 12.16No clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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