Project Overview

• Scientifically driven project clear question
• Not count number of bubbles in system
• Does temperature affect bubble size?
• Take images through appropriate processing steps
• Characterize parameter space and error sources
• Short presentation (5-10 minutes depending on number of students)
• Problem
• Solution
• Key Findings

Sign Up Link

Here you can sign up for the projects.

Our Projects

Performance analysis using sphere packings

We have a series of tomography data obtained acquired using different exposure times to be used as validation reference. The samples is a packing of ball with different diameters.

The task is to implement image processing workflows to

• evaluate different
  • filtering techniques
  • segmentation methods
• evaluate parameterization sensitivity
  • explore how well the individual balls can be identified
  • variations in metrics like volume, and surface area
• the observe the effect of different SNR
• possibly also tomography reconstruction with different number of projections

The goal: to find the ideal combination of processing steps to confidently measure the balls.

Fault evaluation in additive manufacturing

Items produced using additive manufacturing techniques may show faults. In this data set we have a sample which is produced using a combination of 3D printing and cast.

Questions

• where are the faults located in the structures?
• does shape and size depend on the location?
• do the faults cluster or are they evenly distributed?

Detection and Analysis of Calcifications in Breast Mamographies (Z. Wang)

• Calcification count and type is an important criterion for breast cancer diagnosis but it is done manually due to the complexity of breast tissue

Circardian Rhythm in Neuron Cells

Fluorescence signal in neuron cells in the region of the brain responsible for circadian regulation

• Dataset and Kernels on https://www.kaggle.com/kmader/circadian-rhythm-in-the-brain

Questions

• Do the cells move during the experiment?
• How regular is their signal?
• Is the period 24 hours?
• Do nearby cells oscillate together?
• Do they from chunks or groups, over what range do they work?
• Are there networks formed from time-precedence?

Projects of Personal Interest

Running Analysis from Video

The idea would be to quantify body movement in videos of people running (head, legs, arms, hips) in order to start to classify different types of runners and ultimately provide feedback. Ideally the work could be transformed into an app on the phone and used by runners everywhere. The idea could also be applied to different sports and physical events.

• Example Kernel: https://www.kaggle.com/kmader/running-pose-estimate
• Code: https://github.com/ildoonet/tf-pose-estimation
• References: https://github.com/ildoonet/tf-pose-estimation/blob/master/etcs/reference.md

Questions

• Can you quantify the difference between fast and slower runners (maybe use race finisher videos or timelapses from a fixed point on the course: https://www.youtube.com/watch?v=jLcGr_EhUgA)
• Can you identify the difference between amateur and professional racers?
• Can you predict the winner of a race from the first few seconds?
• Can you use anomaly detection to determine runners doing incorrect moves?

Kaggle Projects

Right Whale Identification

• All information copied from here

Currently, only a handful of very experienced researchers can identify individual whales on sight while out on the water. For the majority of researchers, identifying individual whales takes time, making it difficult to effectively target whales for biological samples, acoustic recordings, and necessary health assessments.

Example whale image to identify:

Heart Ejection Fraction from MRI: Data Science Bowl

• All information copied from here

The goal is to automatically measure the ejection fraction of a heart using 4D (3D + time) images of the heart. There are over 17GB of images to analyze and a few hints on the forums on how to use analytical (Fourier method) and Deep Learning approaches.

Connecteomics

• Can a cell network be reconstructed from the fluorescence signals of the firing neurons?
• $3000 in Prize Money for Best Result

Show video

Digit Recognizer

• Can digits be accurately identified from images?
• Automated identification is important for thousands of applications from traffic control to reciept validation.
• Number of solutions already available

(Image borrowed from: Andrej Karpathy, http://karpathy.github.io/2015/05/21/rnn-effectiveness/)