Abstract:
Convolution neural networks (CNNs) have achieved the highest performance today not only for human posture prediction but also for other machine vision tasks (e.g., object...Show MoreMetadata
Abstract:
Convolution neural networks (CNNs) have achieved the highest performance today not only for human posture prediction but also for other machine vision tasks (e.g., object identification, semantic segmentation, images classification). Furthermore, the Attention Module demonstrates their superiority over other conventional networks (AM). As a result, this work focuses on a useful feed-forward AM for CNNs. First, following a stage in the backbone network, feed the feature map into the attention module, which is separated into two dimensions: channel and spatial. The AM then multiplies these two feature maps and passes them on to the next level in the backbone. The network can collect more information in long-distance dependencies (channels) and geographical data, resulting in higher precision efficiency. Our experimental results would also show a difference between the employment of the attention module and current methodologies. As a result of the switch to a High-resolution network (HRNet), the predicted joint heatmap keeps accuracy while reducing the number of parameters compared to the baseline-CNN backbone. In terms of AP, the suggested design outperforms the baseline-HRNet by 2.0 points. Furthermore, the proposed network was trained using the COCO 2017 benchmarks, which are currently available as an open dataset.
Published in: 2022 13th Asian Control Conference (ASCC)
Date of Conference: 04-07 May 2022
Date Added to IEEE Xplore: 20 July 2022
ISBN Information: