<p>IoT devices are a technology recurrent in the contexts of Industry 4.0 and real-time applications. Nonetheless, they suffer from resource limitations, such as processor, RAM, and disc storage. These limitations become more evident when handling demanding applications, such as deep learning, which is well-known for its heavy computational requirements. One way to mitigate processing and storage problems is compressing that deep learning application, reducing its demands. A case in point is robot pose estimation, an application that predicts the critical points of interest on a desired image object. This paper proposes a new CNN for pose estimation while applying the compression techniques of pruning and quantization to reduce demands and improve the response time. While the pruning process reduces the total number of parameters required for inference, quantization decreases the precision of the floating point. We run the approach using a pose estimation task for a robotic arm and compare the results in a high-end and constrained device. As metrics, we consider the number of Floating-point Operations Per Second (FLOPS), the total of mathematical computations, the calculation of parameters, the inference time, and the number of video frames processed per second. In addition, we undertake a qualitative evaluation where we compare the output image predicted for each pruned network with the original one. We reduced the originally proposed network to a 70<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> pruning rate, implying an 88.86<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> reduction in parameters and a 94.45<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> reduction in FLOPS. We reduced the requirement by 70<InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> for the disc storage while increasing error by a mere one perceptual point (P.P). This metric increases input image processing from 11.71 FPS to 41.9 FPS for the Desktop case. Image processing augmented from 2.86 FPS when using the constrained device to 10.04 FPS. The proposed approach’s higher processing rate of image frames allows a much shorter response time.</p>

FCN-Pose: A Pruned and Quantized CNN for Robot Pose Estimation for Constrained Devices

期刊论文
出版模式：

开放获取
发表日期：

2026年1月17日

查看全文

Marrone Silvério Melo Dantas,
Assis T. de Oliveira Filho,
Djamel Fawzi Hadj Sadok,
Judith Kelner,
Ricardo Silva

摘要

IoT devices are a technology recurrent in the contexts of Industry 4.0 and real-time applications. Nonetheless, they suffer from resource limitations, such as processor, RAM, and disc storage. These limitations become more evident when handling demanding applications, such as deep learning, which is well-known for its heavy computational requirements. One way to mitigate processing and storage problems is compressing that deep learning application, reducing its demands. A case in point is robot pose estimation, an application that predicts the critical points of interest on a desired image object. This paper proposes a new CNN for pose estimation while applying the compression techniques of pruning and quantization to reduce demands and improve the response time. While the pruning process reduces the total number of parameters required for inference, quantization decreases the precision of the floating point. We run the approach using a pose estimation task for a robotic arm and compare the results in a high-end and constrained device. As metrics, we consider the number of Floating-point Operations Per Second (FLOPS), the total of mathematical computations, the calculation of parameters, the inference time, and the number of video frames processed per second. In addition, we undertake a qualitative evaluation where we compare the output image predicted for each pruned network with the original one. We reduced the originally proposed network to a 70 $\%$ $%$ pruning rate, implying an 88.86 $\%$ $%$ reduction in parameters and a 94.45 $\%$ $%$ reduction in FLOPS. We reduced the requirement by 70 $\%$ $%$ for the disc storage while increasing error by a mere one perceptual point (P.P). This metric increases input image processing from 11.71 FPS to 41.9 FPS for the Desktop case. Image processing augmented from 2.86 FPS when using the constrained device to 10.04 FPS. The proposed approach’s higher processing rate of image frames allows a much shorter response time.