AI Paper Atlas

Research map / Detection and segmentation / CNN features

SOTR: Segmenting Objects With Transformers

Ruohao Guo, Dantong Niu, Liao Qu, Zhenbo Li

ICCV 2021 · Detection and segmentation › CNN features · approach: Plain and high-resolution ViTs · 127 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
use pure twin transformer instead of original transformer (ablation)SOTR (ResNet-101-FPN) · COCO test-devAP: 37.1 -> 39.7Improved
add two 3x3 convolutions to the twin attention module (hybrid twin) (ablation)SOTR pure twin transformer · COCO test-devAP: 39.7 -> 40.2Improved
increase original transformer depth from 6 to 12 (ablation)SOTR original transformer · COCO test-devAP: 36.2 -> 37.1Improved
increase twin transformer depth from 6 to 12 (ablation)SOTR hybrid twin · COCO test-devAP: 37.6 -> 40.2Improved
use twin transformer instead of original at depth 12 (memory) (ablation)SOTR · COCO test-devmemory: 10696M -> 5492MImproved
substitute P5 of FPN with transformer features in the upsampling module (ablation)SOTR · COCO test-devAP: 38.8 -> 40.2Improved
also substitute P4 with transformer features (ablation)SOTR · COCO test-devAP: 40.2 -> 39.9; APL 73.0 -> 73.7Mixed
use dynamic convolution kernels instead of static kernels (ablation)SOTR with twin transformer · COCO test-devAP: 38.6 -> 40.2Improved
add twin transformer to model with dynamic kernels (ablation)SOTR · COCO test-devAP: 39.7 -> 40.2Improved
use ResNet-101-FPN instead of ResNet-50-FPN backbone (ablation)SOTR · COCO test-devAP: 37.5 -> 40.2Improved
compare SOTR against Mask R-CNN with the same ResNet-101-FPN backboneSOTR vs Mask R-CNN · COCO test-devAP: 35.7 -> 40.2 (+2.4% AP)Improved
use SOTR-RT instead of YOLACT-550 for real-time instance segmentationSOTR-RT-736 · COCO test-devAP: 28.2 -> 30.7; FPS 44.1 -> 25.2Mixed

Later papers that compared with this method

1 later accepted paper reports SOTR as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 8 of 9 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