AI Paper Atlas

Research map / Video / Action recognition

Uni-AdaFocus: Spatial-Temporal Dynamic Computation for Video Recognition

Yulin Wang, Haoji Zhang, Yang Yue, Shiji Song, Chao Deng, Junlan Feng, Gao Ming Huang

TPAMI 2024 · Video › Action recognition · approach: Contrastive video representation learning · 20 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
train end-to-end (AdaFocusV2) instead of three-stage RL (AdaFocusV1)AdaFocus · ActivityNet (patch 128^2)mAP: 75.0% -> 79.0%Improved
train end-to-end instead of three-stage RLAdaFocus · ActivityNet (patch 128^2)training wall-time: 7.2h -> 3.7hImproved
add Uni-AdaFocus refinements (feature-interpolation policy training, deformable patches, no sample-wise) over AdaFocusV2AdaFocusV2 · ActivityNet (training efficiency comparison)mAP: improves mAP by 1.6-1.9%Improved
add conditional-exit (sample-wise dynamic computation) (ablation)Uni-AdaFocus (128^2) · ActivityNetmAP and GFLOPs: 80.7% -> 80.4%, 27.2 -> 18.1 GFLOPsImproved
add conditional-exit at a lower budget (ablation)Uni-AdaFocus (128^2) · ActivityNetmAP and GFLOPs: 80.7% -> 78.3%, 27.2 -> 9.1 GFLOPsImproved
add conditional-exit (ablation)Uni-AdaFocus (128^2) · FCVIDmAP and GFLOPs: 86.4% -> 86.2%, 27.2 -> 18.1 GFLOPsImproved
add conditional-exit (ablation)Uni-AdaFocus (128^2) · Mini-KineticsTop-1 accuracy and GFLOPs: 75.8% -> 75.7%, 27.2 -> 18.1 GFLOPsImproved
use Uni-AdaFocus instead of AdaFocusV3 at inferenceUni-AdaFocus · ActivityNetinference cost: 2.3-2.8x lower costImproved
use Uni-AdaFocus instead of OCSampler to reach the same mAPUni-AdaFocus · ActivityNetGFLOPs per video: approximately 4.8x lessImproved
use Uni-AdaFocus (128^2) instead of OCSamplerUni-AdaFocus · ActivityNetmAP: 76.9% -> 80.4%, 21.7 -> 18.1 GFLOPsImproved
use Uni-AdaFocus (128^2) instead of NSNetUni-AdaFocus · FCVID and Mini-KineticsmAP / Top-1: 83.9% -> 86.2%, 73.6% -> 75.7%Improved
add Uni-AdaFocus on top of TSM+ backbone (96^2, w/o sample-wise)TSM+ · Something-Something V1Top-1 accuracy and GFLOPs: 47.0% -> 47.5%, 35.1 -> 8.8 GFLOPsImproved

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