AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Inverted Pyramid Multi-task Transformer for Dense Scene Understanding

Hanrong Ye, Dan Xu

ECCV 2022 · Architectures and efficiency › Attention and transformers · approach: Efficient image transformers · 95 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 UP-Transformer block (UTB) (ablation)InvPT Baseline (MT) · NYUD-v2 SemsegmIoU: 41.06 -> 43.18Improved
add UP-Transformer block (UTB) (ablation)InvPT Baseline (MT) · NYUD-v2 DepthRMSE: 0.6350 -> 0.5643Improved
add UP-Transformer block (UTB) (ablation)InvPT Baseline (MT) · NYUD-v2 BoundaryodsF: 76.00 -> 76.10No clear effect
add cross-scale attention message passing (AMP) (ablation)InvPT w/ UTB · NYUD-v2 SemsegmIoU: 43.18 -> 43.64Improved
add cross-scale attention message passing (AMP) (ablation)InvPT w/ UTB · PASCAL-Context SemsegmIoU: 72.34 -> 73.29Improved
add encoder feature aggregation (EFA) (ablation)InvPT w/ UTB+AMP · NYUD-v2 SemsegmIoU: 43.64 -> 44.27Improved
add encoder feature aggregation (EFA) (ablation)InvPT w/ UTB+AMP · PASCAL-Context ParsingmIoU: 61.78 -> 62.73Improved
add encoder feature aggregation (EFA) (ablation)InvPT w/ UTB+AMP · NYUD-v2 BoundaryodsF: 76.10 -> 76.10No clear effect
train all tasks jointly with a plain multi-task decoder instead of single-taskInvPT Baseline · NYUD-v2 SemsegmIoU: 43.29 -> 41.06Hurt
use full InvPT multi-task model instead of single-task baselineInvPT · NYUD-v2MT performance Delta_m: +2.59Improved
use full InvPT multi-task model instead of single-task baselineInvPT · PASCAL-ContextMT performance Delta_m: +1.76Improved
scale encoder from Swin-T to Swin-BInvPT · NYUD-v2 SemsegmIoU: 44.27 -> 50.97Improved

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