Análise de desempenho de mecanismos de atenção para estimativa de pose 2D baseada em resnet-50

Abstract

2Dhumanpose estimation is a fundamental computer vision problem focused on locating human anatomical keypoints. While deep learning, particularly Convolutional Neural Networks (CNNs), has driven significant progress, attention mechanisms have emerged as an effective method to enhance a model’s focus on salient image regions. This dissertation presents a comparative study analyzing the impact of six different attention mechanisms on 2D human pose estimation when integrated into a ResNet-50-based CNN baseline. The evaluated mechanisms are: Convolutional Block Attention Module (CBAM), Coordinate Attention, Global Context Attention, Self-Attention, Multi-Head Attention, and SimAM (Simple, Parameter-Free Attention Module). All models were trained and evaluated on the MS COCO dataset under a unified experimental protocol. Quantitative results show that all attention mechanisms improved performance over the baseline. Coordinate Attention and CBAM were the most effective, achieving an Average Precision (AP) of 67.7% (+1.5 p.p.) and 67.6% (+1.4 p.p.), respectively, with CBAM also leading in the AP75 metric. A cost-benefit analysis confirmed these two models offered the best performance gains with a minimal increase in parameters and FLOPS. Conversely, the computationally expensive Self-Attention yielded one of the smallest gains, while the parameter-free SimAM offered the lowest improvement at no extra cost. In conclusion, this work demonstrates that integrating attention mechanisms is an effective strategy for human pose estimation. Specifically, approaches emphasizing explicit spatial information, like CBAM and Coordinate Attention, provide an excellent balance between performance improvement and computational efficiency