Chat with Michael Holroyd on Computer Vision

[Chang Lee]

Participants

• Michael Holroyd
• Chang Lee

Background

I learned that Michael had experience in computer vision. Being neck deep in building computer vision software now, I set up a penny chat with him to learn from his experience.

Michael has a PhD in Computervision from UVA and cofounded a company after graduate school. These days he's working on https://scenethink.com/

His research in graduate school was on photogrammetry, in particular, how to generate 3D models from datasets, which he called the "old school computer vision."

Takeaways

CV is a field with lots of pie-in-the-sky startups. He thinks it is going through the phase where people started to solve more business problem rather than pitching crazy ideas.

I asked him what is an example of crazy idea. He said one example of impossible startup idea is measuring foot with iPhone and recommend what kind of shoes to wear.

This sounds normal to me, so I asked him why he thinks it is an extremely hard problem. He gave me a few reasons: 

• Foot is constantly moving - not being able to keep a person's foot still means it's a difficult task for photogrammetry.
• photogrammetry is mostly through triangulation. 
• Lack of defining feature: on the camera, foot pixels look like a blob and there's no feature to anchor on.
• Subsurface scattering: foot is not opaque and has severe subsurface scattering 
• Need to build a deformable model into the process.

He mentioned a new feature on the new iPad called time-of-flight sensor and he thinks it is a much better solution than triangulation. For example, Xbox Kinect 1 is based on triangulation while Kinect 2 is based on time-of-flight. The current implementation merges both.

We also talked about activity and gesture recognition. It is a place where I can tap into for features unique to videos: sometimes it can be hard to answer what a person is doing in a still frame.

Following computer vision research

He gets his information through two conferences

• SIGGRAPH: the best
• CVPR: next best

He reads through the titles and abstracts and sometimes watch the videos included in them to see if interesting new research comes up.

Misc 

• Prodigy: an interesting labeling tool from the makers of spaCy