AI Paper Atlas

Research map / Video / Action recognition

MGSampler: An Explainable Sampling Strategy for Video Action Recognition

Yuan Zhi, Zhan Tong, Limin Wang, Gangshan Wu

ICCV 2021 · Video › Action recognition · approach: Disentangled action recognition · 89 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
replace TSN uniform sampling with motion-guided sampling (feature difference)TSM ResNet50, 8 frames · Something-Something V2top-1 accuracy: 57.9 -> 60.1Improved
replace TSN uniform sampling with motion-guided sampling (feature difference)TSM ResNet50, 8 frames · Something-Something V1top-1 accuracy: 45.6 -> 47.3Improved
replace TSN sampling with motion-guided sampling (feature difference)TSM ResNet50, 8 frames · Diving48top-1 accuracy: 35.2 -> 37.4Improved
replace TSN sampling with motion-guided sampling (feature difference)TSM ResNet50, 8 frames · UCF101top-1 accuracy: 94.5 -> 95.2Improved
replace TSN sampling with motion-guided sampling (feature difference)TSM ResNet50, 8 frames · HMDB51top-1 accuracy: 72.6 -> 73.8Improved
replace TSN sampling with motion-guided sampling (feature difference)TSM ResNet50, 8 frames · Jestertop-1 accuracy: 96.5 -> 97.5Improved
use feature-level difference instead of image-level difference as motion representation (ablation)MGSampler · Something-Something V2top-1 accuracy: 59.8 -> 60.1Improved
set smoothing mu=0.5 instead of mu=1 (original motion magnitude) (ablation)MGSampler image diff · Something-Something V1top-1 accuracy: 45.7 -> 47.1Improved
set smoothing mu=2 instead of mu=0.5 (exaggerate motion differences) (ablation)MGSampler image diff · Something-Something V1top-1 accuracy: 47.1 -> 42.8Hurt
use cumulative-motion-distribution sampling instead of motion-magnitude-only sampling (ablation)MGSampler · Something-Something V1top-1 accuracy: 41.5 -> 47.3Improved
use motion-guided sampling instead of fixed stride samplingTSM ResNet50 · Something-Something V2top-1 accuracy: 53.4 -> 60.1Improved
add MGSampler to TSM ResNet101 with 8 framesTSM · Something-Something V1top-1 accuracy: 46.9 -> 47.8Improved

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