Project P3:
Identifying Cell Components

Project Goal

The Department of Molecular Pharmacology has a database of thousands of microscope photographs of individual cells. The Department is interested in using computer vision to identify specific components of these cells -- the corpus is too large to annotate by hand. Training a classifier on a corpus of several hundred photographs is a great way to start. However, the cell components are sometimes occluded. In addition, the components really exist in 3D so they may look quite different depending on their orientation when projected into the 2D photograph. Perhaps a sequence of classifiers will be required. Furthermore, there may be geometry cues like the proximity of certain cell components to other components. Ultimately, this is a very challenging object identification problem where the domain is really quite unconstrained. [I will need to meet with them again before I can really say more.]

Project Scope

The main deliverable for this project is a program that can read in an image of a cell and annotate the cell's components. This project is a new collaboration between the Stanford Artificial Intelligence Laboratory and the Department of Molecular Pharmacology in the Medical School. The scope may evolve as the project does.

Tasks

• Acquire Cell Input Images / Meet with Molecular Pharmacology
• Find features using the Shi-Tomasi Algorithm in OpenCV (1 week)
• Try Simple Ideas (Color, Texture, Etc.) as a Classifier (1-2 weeks)
• Train a Haar Classifier (2-3 weeks)
• --MIDTERM REPORT--
• Build a More Advanced Classifier using 3D->2D Projection (2-3 weeks)
• Use Higher-Level Knowledge in your Classifier (2-3 weeks)

Project Status

Guillem Pratx (pratx at stanford dot edu),
Pierre Yves Laligand (laligand at stanford dot edu),
Joelle Barral (jbarral at stanford dot edu).

Point of Contact

David Stavens, Dan Morris

Midterm Report

not yet submited

Final Report

not yet submitted