University of Twente Student Theses
As of Friday, 8 August 2025, the current Student Theses repository is no longer available for thesis uploads. A new Student Theses repository will be available starting Friday, 15 August 2025.
Optimising Deep Learning Models for Human Pose Estimation
Aswani Ahuja, Abhishek (2024) Optimising Deep Learning Models for Human Pose Estimation.
PDF
743kB |
Abstract: | This thesis proposes to optimize the TransPose-R model, a transformer-based architecture for human pose estimation. The primary goal is to reduce the model size while maintaining acceptable performance, addressing the need for efficient, deployable pose estimation systems in resource-constrained environments. The research plans to explore optimization techniques Network Slimming. This involves pruning channels from the CNN backbone by using scaling factors to identify less important channels, and then fine-tuning the pruned model. The study will investigate different pruning strategies, varying in sparsity factors and fine-tuning approaches. Experiments will be conducted using the COCO dataset, with model effectiveness measured by validation accuracy and compression ratio. This research aims to contribute to the field of efficient deep learning for computer vision tasks, particularly in human pose estimation. It seeks to demonstrate the potential for creating more compact models without significant performance loss, exploring the trade-offs between model size and accuracy. The project's outcomes have implications for deploying advanced pose estimation systems on resource-constrained devices, potentially enabling more widespread use of real-time human pose estimation across various applications. |
Item Type: | Essay (Master) |
Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
Subject: | 54 computer science |
Programme: | Computer Science MSc (60300) |
Link to this item: | https://purl.utwente.nl/essays/103252 |
Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Repository Staff Only: item control page