If you are interested in working on one of the projects listed below, please send an email to Prof. Marcel Salathe in which you answer the following questions:
- Which project(s) are you interested in?
- Are you looking for a thesis or a semester project / Lab immersion? Priority will be given to Master Thesis Projects (full-time)
- Why did these specific projects catch your attention?
- How much time could you spend on the project per week? How many classes are you planning on taking in addition to the project?
Please attach your CV and links to projects you developed (e.g. GitHub repo)
Background:
The accurate measurement of food volume is critically important in nutritional science and food technology. It’s essential for tasks like dietary assessment, portion size measurement, and monitoring caloric intake. Traditional volume estimation methods, which often involve manual measurements or basic assumptions, may not accurately represent the diverse shapes and sizes of various food items.Project Description:
This project is designed to utilize deep learning for estimating the volume of food items from images. The project utilizes a dataset comprising 100,000 images of different food items. Each image is meticulously annotated, detailing the segmentation of each food item, classification of these segments, and the estimated weight of each item. The main goal is to develop a model that precisely estimates food volume, using this rich dataset and advanced deep learning techniques.Objectives:
Model Development:Explore and apply state-of-the-art deep learning models, while considering using multi-modal inputs, combining image data with meta-data (classification labels and weights) for more accurate volume estimation.Volume Estimation Strategy:Formulate a method to calculate the volume of food items, either directly from segmentation masks or indirectly by correlating known weights with estimated volumes.
Use geometric and spatial analysis to refine volume estimation, taking into account the shape and orientation of food items in the images.
Model Training and Evaluation:Train the model with suitable loss functions and optimization methods, focusing on accurate volume estimation.
Assess the model’s accuracy by comparing predicted volumes with manually measured volumes for a portion of the dataset.
Results Analysis and Optimization : Evaluate the results to determine the model’s strengths and weaknesses.Improve the model architecture, training procedure, and volume estimation approach based on this evaluation, aiming to increase the precision of the volume estimations.Expected Outcomes:
The successful completion of this project is anticipated to produce a robust deep learning model that can precisely estimate the volume of various food items from images. This advancement promises to significantly enhance areas such as nutritional science, dietary management, and food technology, offering a scalable and automated approach to food volume estimation. Moreover, the insights from this project may open doors to further research and applications in related fields like automated dietary tracking and smart food preparation systems.