University of Twente Student Theses


Pose estimation with deep neural network trained on tiny datasets

Liang, Y. (2022) Pose estimation with deep neural network trained on tiny datasets.

[img] PDF
Abstract:This report presents my project on the pose estimation from training a deep neural network with small datasets. The primary object of this project comes from the project on a soft medical robot. Since soft robots have unlimited degrees of freedom and convenience, being hard to control is a great challenge. A camera and a sensor are attached to it to have enough information to control the robot’s end- effector. The camera can shoot videos while the robot is working, and the sensor provides the pose information of the end-effector. However, the implementation of the sensor is not only complicated but also expensive. Fortunately, the camera pose collected by the sensor can now be estimated with deep learning algorithms. Deep neural networks usually perform much better on large datasets. What comes with the great advantage of the performance of deep neural networks is the constraints on the dataset size. A pruning strategy on the most state-of-the-art deep learning frameworks is proposed to overcome the limited performance while training with small datasets, which is more common in practical use. According to experimental results, translation and rotation test errors are reduced after training a model on small datasets. Therefore, the strategy could also be promoted to other applications when the training dataset is of limited size.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Electrical Engineering MSc (60353)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page