AI Paper Atlas

Research map / Graph learning / Knowledge and temporal graphs

Ske2Grid: Skeleton-to-Grid Representation Learning for Action Recognition

Dongqi Cai, Yangyuxuan Kang, Anbang Yao, Yurong Chen

ICML 2023 · Graph learning › Knowledge and temporal graphs · approach: Graph reasoning for visual recognition · 2 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 GIT+UPT+PLS grid representation (Ske2Grid_8x8) to ST-GCNST-GCN* on 2D estimated skeletons · six datasets (2D est.)Top-1 Acc (%): +1.12 to +5.79 across datasetsImproved
add UPT to GIT (grid 5x5) (ablation)ST-GCN* grid representation · NTU-60 XSub (2D est.)Top-1 Acc (%): +1.50 (84.70 to 86.20)Improved
add UPT to GIT (grid 8x8) (ablation)ST-GCN* grid representation · NTU-60 XSub (2D est.)Top-1 Acc (%): +0.97 (84.81 to 85.81)Improved
add PLS to GIT (grid 8x8) (ablation)ST-GCN* grid representation · NTU-60 XSub (2D est.)Top-1 Acc (%): +1.30 (84.81 to 86.11)Improved
add both UPT and PLS to GIT (grid 8x8) (ablation)ST-GCN* grid representation · NTU-60 XSub (2D est.)Top-1 Acc (%): +3.74 (84.81 to 88.55)Improved
add UPT alone to GIT (grid 9x9) (ablation)ST-GCN* grid representation · NTU-60 XSub (2D est.)Top-1 Acc (%): -0.85 (84.10 to 84.95, GIT only at 9x9 is 84.10 so UPT helps; but GIT+UPT 84.95 falls below ST-GCN 85.25)Hurt
Ske2Grid convolution vs ST-GCN convolution on learnable grid (D5x5) (ablation)grid patch with learnable UPT and GIT · NTU-60 XSub (2D est.)Top-1 Acc (%): +5.81 (80.39 to 86.20)Improved
Ske2Grid convolution vs ST-GCN convolution on fixed grid (D5x5) (ablation)grid patch with fixed UPT and GIT · NTU-60 XSub (2D est.)Top-1 Acc (%): ST-GCN fixed 84.34 vs ST-GCN* learnable 86.20 (fixed-grid Ske2Grid conv not reported separately)Mixed
apply Ske2Grid convolution to CTR-GCN backboneCTR-GCN on 3D GT skeletons · NTU-60 XSub / XView, NTU-120 XSub / XSet (3D GT)Top-1 Acc (%): +0.46 (NTU-60 XSub), +0.26 (XView), +0.22 (NTU-120 XSub), +0.09 (XSet)Improved
apply Ske2Grid to ST-GCN on 3D GT skeletons (Ske2Grid_9x9)ST-GCN on 3D GT skeletons · NTU-60 and NTU-120 (3D GT)Top-1 Acc (%): +0.63 (NTU-60 XSub), +0.64 (XView), +1.60 (NTU-120 XSub), +1.44 (XSet)Improved
multi-scale ensemble (D5x5 to D8x8) with joint, bone, joint-motion, bone-motion modalitiesSke2Grid on 2D estimated skeletons · NTU-60 XSub / XView, NTU-120 XSub / XSetTop-1 Acc (%): NTU-60 XSub 93.8 and XView 98.6, best on three of four benchmarksImproved
use squared grid patches instead of rectangular ones (5x5 vs 4x6, 5x6, 5x7) (ablation)Ske2Grid grid patch shape · NTU-60 XSub (2D est.)Top-1 Acc (%): 86.20 (5x5) vs 85.91 (5x6), 85.61 (4x6), 84.25 (5x7)Improved

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