Enhancing Segmentation Accuracy: The Role of Photorealism in Transfer Learning from Synthetic to Real Point Clouds
Supervisor: Katherine James
What you will be doing
A crucial step in deep learning for plant studies is the segmentation of a plant into its respective organs – leaves, fruit, stems, etc. However, the state-of-the-art for training segmentation models requires annotated data, which is time-consuming to create.
In this project, we aim to use synthetically generated plants to explore how realism impacts performance when using these as a basis for pre-training segmentation networks.
This project provides the opportunity to gain exposure to:
- Working with real 3D data (https://lcas.github.io/LAST-Straw/)
- 3D annotation in Segments.ai
- Generating synthetic data by building upon existing strawberry models
- State-of-the-art computer vision techniques
Figure 1: Synthetically generated strawberry plants demonstrating different patterns of fruit display
What skills would be useful to have for this project
A successful candidate will have a good knowledge of Python and a willingness to learn and work independently. Knowledge of machine learning frameworks (eg. PyTorch) would be beneficial. Knowledge of L-Py is not expected but would be a bonus.
How to get more information and apply
For more information or to apply, please contact Katherine (kajames@lincoln.ac.uk) by email in the first instance. Your application should include a brief description of your interests and your suitability for this project, as well as a CV. Candidates will be selected based on the content of the application, followed by a brief chat on Teams/in-person to confirm your suitability.
The application deadline is 28 June, and the successful candidate will be notified by email by the 5th of July. The start date would be 29 July (negotiable).