Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Coversion of tomato dataset to different lighting conditions
We have a large dataset of tomato plant pictures in greenhouses with massive annotations of different objects. The goal is to convert images to other lighting conditions using deep neural networks. This includes changing light direction, intensity, and also cloudy or night conditions with artificial illumination.
Keywords: Deep Learning, Vision, Neural Networks
At RSL, we have robots that manipulate objects in real-world environments. In the greenhouse project (www.floatingrobotics.com), we aim to harvest and de-leaf tomato plants in glass greenhouses. Our robots need to perceive and understand their surroundings as well as objects of interest. To maximize efficiency and robustness, this capability needs to extend in different lighting conditions. Current neural networks are trained to tolerate variations to some extent but need a much more diverse dataset to achieve maximum robustness. We already have a large annotated dataset, yet it's taken at a certain time of the day. Collecting images under other lighting conditions is feasible, though annotating them takes too much time. Repeating this for different tomato varieties at different stages of growth renders this task nearly impossible. In this project, we want to train neural networks that augment the original dataset by changing light direction and intensity. We want to go one step further and convert them to night conditions where the robot illuminates the scene with its LEDs. For this purpose, we will take videos of the rows by mounting the robot on a pipe trolley and moving at a constant speed. Repeating this for one day will generate movies with different lighting conditions and minimal scene change. Synchronization of videos together would then produce large datasets for training desired neural networks. The concrete goal of the project within its timeframe is to make 2 converters for morning to noon, and noon to night conditions.
Project Co-supervisor: Dr. Mohsen Moosavi https://www.imperial.ac.uk/people/seyed.moosavi
At RSL, we have robots that manipulate objects in real-world environments. In the greenhouse project (www.floatingrobotics.com), we aim to harvest and de-leaf tomato plants in glass greenhouses. Our robots need to perceive and understand their surroundings as well as objects of interest. To maximize efficiency and robustness, this capability needs to extend in different lighting conditions. Current neural networks are trained to tolerate variations to some extent but need a much more diverse dataset to achieve maximum robustness. We already have a large annotated dataset, yet it's taken at a certain time of the day. Collecting images under other lighting conditions is feasible, though annotating them takes too much time. Repeating this for different tomato varieties at different stages of growth renders this task nearly impossible. In this project, we want to train neural networks that augment the original dataset by changing light direction and intensity. We want to go one step further and convert them to night conditions where the robot illuminates the scene with its LEDs. For this purpose, we will take videos of the rows by mounting the robot on a pipe trolley and moving at a constant speed. Repeating this for one day will generate movies with different lighting conditions and minimal scene change. Synchronization of videos together would then produce large datasets for training desired neural networks. The concrete goal of the project within its timeframe is to make 2 converters for morning to noon, and noon to night conditions.
Project Co-supervisor: Dr. Mohsen Moosavi https://www.imperial.ac.uk/people/seyed.moosavi
1- Literature study
2- Data collection and synchronization
3- Software programming
4- Test on the real robot
1- Literature study 2- Data collection and synchronization 3- Software programming 4- Test on the real robot
- Strong record of courses and/or projects in machine vision and deep learning.
- Python Programming skills
- Strong record of courses and/or projects in machine vision and deep learning. - Python Programming skills