AI Paper Atlas

Research map / Video / Action recognition

Gate-Shift-Fuse for Video Action Recognition

Swathikiran Sudhakaran, Sérgio Escalera, Oswald Lanz

TPAMI 2023 · Video › Action recognition · approach: Spatiotemporal CNNs for action recognition · 38 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 learnable channel fusion in place of sum fusion (ablation)GSF (BNInception) · Something-Something V1 (BNInception, 8 frames)accuracy: 47.24 -> 48.44Improved
make spatial gating learnable instead of fixed to 1 (TSM-like) (ablation)GSF (BNInception) · Something-Something V1 (BNInception, 8 frames)accuracy: 45.57 -> 48.44Improved
make gating learnable instead of fixed to -1 (differencing) (ablation)GSF (BNInception) · Something-Something V1 (BNInception, 8 frames)accuracy: 45.89 -> 48.44Improved
add gating (non-zero) over fixed gating 0 (TSN) (ablation)GSF (BNInception) · Something-Something V1 (BNInception, 8 frames)accuracy: 17.25 -> 48.44Improved
add GSF to all 10 Inception blocks instead of last block only (ablation)TSN BNInception · Something-Something V1 (BNInception, 8 frames)accuracy: 31.9 -> 48.44Improved
add GSF to 5 blocks instead of 1 (ablation)TSN BNInception · Something-Something V1 (BNInception, 8 frames)accuracy: 31.9 -> 45.24Improved
plug GSF into the least-convolution Inception branch rather than all branches (ablation)GSF (BNInception) · Something-Something V1 (BNInception, 8 frames)accuracy: 46.55 -> 48.44Improved
shift all channels instead of 12.5% (ablation)GSF (BNInception) · Something-Something V1 (BNInception, 8 frames)accuracy: 45.62 -> 48.44Improved
shift only 25% of channels instead of 100% (ablation)GSF (ResNet-50) · Something-Something V1 (ResNet-50, 8 frames)accuracy: 47.63 -> 50.17Improved
add GSF to 16 bottleneck layers instead of 4 (ablation)TSN ResNet-50 · Something-Something V1 (ResNet-50, 8 frames)accuracy: 45.97 -> 50.17Improved
apply GSF after conv2 vs after conv3 of ResNet bottleneck (ablation)GSF (ResNet-50) · Something-Something V1 (ResNet-50, 8 frames)accuracy: 47.15 -> 47.63Mixed
add MotionSqueeze module to GSF ResNet-50 (16 frames, 3x2 clips)GSF ResNet-50 · Something-Something V1accuracy: 52.33 -> 54.04Improved

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