Resource-Aware Video Action Recognition Through Adaptive Temporal Sampling: An Efficiency–Accuracy Benchmarking Framework

<title>Abstract</title> <p>Human action recognition (HAR) systems deployed in resource-constrained environments must balance recognition performance and computational efficiency. This study presents a resource-aware benchmarking framework based on adaptive temporal sampling to investigate efficiency–accuracy trade-offs in video action recognition. The proposed framework integrates adaptive temporal resolution selection, virtual battery modelling, kernel-wise resource estimation, and resource-aware optimization to evaluate the behaviour of representative spatiotemporal architectures under predefined computational budgets. Experiments were conducted using R3D-18, MC3-18, and R(2 + 1)D-18 architectures on the UCF101 and HMDB51 benchmark datasets under multiple energy-budget settings. Results demonstrate that adaptive temporal sampling maintains stable recognition behaviour across varying computational constraints while reducing redundant temporal processing. Among the evaluated architectures, R(2 + 1)D-18 achieved substantially lower estimated energy consumption, requiring approximately 74–76% less energy than R3D-18 under comparable evaluation conditions. Analysis of controller decisions further revealed that Full and Half temporal sampling strategies accounted for the majority of selections, whereas Quarter sampling was not selected under any evaluated configuration, indicating that moderate temporal sparsification provides favourable efficiency–accuracy trade-offs without excessive information loss. The proposed framework provides a reproducible methodology for analysing the relationship between temporal resolution, computational cost, and recognition performance, and offers practical insights for the design of resource-aware video action recognition systems intended for deployment in computationally constrained environments.</p>