Revisiting Attention Modules for Still-Image Human Action Recognition: A Multi-Seed Evaluation on Stanford-40

Attention modules such as Squeeze-and-Excitation are often reported to improve convolutional neural networks for human action recognition in still images. Most of these reports, however, rely on a single training run, and published scores on the same benchmark are frequently obtained under different train–test splits. This study examines whether the reported benefit remains under a controlled evaluation. Four modules, namely Squeeze-and-Excitation (SE), the Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA), and the parameter-free SimAM, were inserted into ResNet-34 at four positions and trained on the Stanford-40 Actions dataset under its official split with one fixed training recipe, giving 44 trained models. The four all-stage configurations were repeated with six random seeds and compared using a paired Wilcoxon signed-rank test and an exact McNemar test. Replacing the official split with a random 80/20 split increased test accuracy by 3.19 to 4.52 percentage points across four backbones, a larger change than any module produced. Across six seeds, SE changed macro-F1 by +0.06 points, within the seed-to-seed range of the unmodified baseline, whereas ECA and SimAM reduced it by 1.56 and 0.74 points. No module improved on the baseline by a statistically supported margin, yet every module increased single-image CPU latency by 4.3% to 18.3% while adding less than 0.02% in floating-point operations. The findings show that the evaluation protocol and run-to-run variance affect reported scores more than the modules themselves, and a reporting checklist is proposed for future evaluations of such modules