AI Paper Atlas

Research map / Detection and segmentation / CNN features

Multi-Exit Semantic Segmentation Networks

Alexandros Kouris, Stylianos I. Venieris, Stefanos Laskaridis, Nicholas Lane

ECCV 2022 · Detection and segmentation › CNN features · approach: Context modeling for semantic segmentation

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 exit-aware (exit-dropout) pre-training and frozen-backbone stage vs Frozen (ablation)MESS DRN-50 uniform exits · MS COCO valaccuracy across exits: up to 12.57ppImproved
use two-stage training instead of end-to-end joint training (ablation)MESS DRN-50 uniform exits · MS COCO valaccuracy across exits: up to 3.38 ppImproved
train exits end-to-end jointly (E2E) rather than vanilla single-exit (ablation)DRN-50 backbone · MS COCO valfinal-exit accuracy: 1.57 pp accuracy drop in the final exitHurt
add Positive Filtering Distillation instead of CE (ablation)MESS second stage · MS COCO valaccuracy: up to 1.8 ppImproved
add Positive Filtering Distillation instead of KD (ablation)MESS second stage · MS COCO valaccuracy: up to 2.32 ppImproved
add Positive Filtering Distillation instead of SelfDistill (ablation)MESS second stage · MS COCO valaccuracy: up to 1.28 ppImproved
use image-level confidence metric instead of averaging per-pixel confidence (ablation)MESS exit policy · MS COCO valaccuracy at equal speed: up to 6.34 pp (1.17 pp on average)Improved
use tailored searched exit architectures instead of uniform FCN heads (ablation)MESS · MS COCO valmIoU under latency budget: up to 19.3 pp (10.9 pp on avg.)Improved
use tailored searched exit architectures instead of uniform FCN heads (ablation)MESS · MS COCO valFLOPs under accuracy target: up to 3x (2.4x on avg.)Improved
apply MESS to DRN-50 for latencyDRN-50 · MS COCO valworkload: 3.36x workload reduction, 2.23x latency speedup, no accuracy dropImproved
allow <=1 pp accuracy degradation in MESS search (ablation)MESS DRN-50 · MS COCO valworkload: 4.01x workload, 2.65x latencyImproved
search for accuracy under DRN workload budgetMESS DRN-50 · MS COCO valmIoU: +5.33 pp mIoU with 1.22x fewer GFLOPsImproved

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