AI Paper Atlas

Research map / Detection and segmentation / CNN features

AutoFocus: Efficient Multi-Scale Inference

Mahyar Najibi, Bharat Singh, Larry S. Davis

ICCV 2019 · Detection and segmentation › CNN features · approach: Efficient DETR query design · 147 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
apply AutoFocus FocusChips at finer scales instead of full 3-scale pyramidSNIPER detector · COCO test-devAP: 47.9 -> 47.9 at 1175^2 pixels vs 1910^2No clear effect
use AutoFocus for faster inferenceSNIPER detector · COCO test-devimages per second: 2.5 -> 6.4 images per secondImproved
reduce pixels processed to 930^2AutoFocus · COCO test-devAP: 47.9 -> 47.2Hurt
reduce pixels processed to 860^2AutoFocus · COCO test-devAP: 47.9 -> 46.9Hurt
add scale 3 (highest resolution) to the lower two scales (ablation)multi-scale SNIPER · COCO val-2017AP: 45.4 -> 47.5Improved
add FocusChip processing of one fifth of scale 3 to scales 1 and 2 (ablation)multi-scale SNIPER · COCO val-2017AP: 45 -> 47.5Improved
reduce pixels at highest resolution by 11 times with FocusChips (ablation)AutoFocus · COCO val-2017AP: 0.5% drop in mAPHurt
reduce overall pixels processed by 5 times (ablation)AutoFocus · COCO val-2017AP: 1% drop in mAPHurt
reduce pixels processed by 2.8 times (ablation)AutoFocus · COCO val-2017AP: 2.8 times fewer pixels without any loss of performanceNo clear effect
use COCO-tuned hyper-parameters on PASCAL VOC (AutoFocus*) instead of SNIPER full processingAutoFocus · PASCAL VOC 2007 testAP70: 80.5 -> 79.5Hurt
use AutoFocus instead of Deformable ConvNet v2 at similar processed areadetector · PASCAL VOC 2007 testAP70: 4.6% better AP at 0.7 IoUImproved
reduce minimum chip size from 512 to 64 (ablation)FocusChip generation · COCOtheoretical speedup: theoretical speedup of ~10 times but detection performance can dropMixed

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