My experience was one that contained some regrets, and as I’ve talked to other Ph.D.’s, there is always something they regretted in their degree. No matter how many papers were published, no matter the citations, no matter the results, they could have done something better. Honestly, I felt like an imposter after graduating. I had a huge expectation of what I would do that was not quite in line with reality. However, I share these things in the hope that they may help someone else deal with post-Ph.D. hangover.
There are four main components of my dissertation work I learned great lessons from. I initially wanted to write that I would change these things, but on reflection, these experiences provided me more than everything going super amazing and successful. They gave me real world experiences that research is hard in that it requires a cultivation of patience.
I built a 3D face scanner that a person would walk through to have their face scanned. It required a person to move because the sensor itself was stationary cameras. I had a laser light screen to provide structured light, and the person’s movement would be in place of the sensor usually moving. This required cameras that ran at 100 fps. So I got cameras that run at 60 fps but could run at 100 fps with a reduced Region of Interest (ROI). Remember: This was 2007.
I started with a single day, small data collection, and then I planned for larger. I did another slightly larger collection with the aim for much larger scale. I started collecting data in the fall of 2008, and I noticed some gaps in the 3D faces. I found these gaps were in all of the 3D images just in different places. It turns out that the camera or the operating system was pausing for 200ms every 2 seconds. This means, I would lose 20 frames every 2 seconds.
The three cameras were running on two fire cards to a linux box running Ubuntu. I tried a single card, a single camera, 60 fps, and 30 fps to no avail. I consulted with everyone in the know about operating systems in the department, and I scoured the internet. Alas, there was no answer after a month of trying. I even wrote a new code to use a different protocol to pull the images, but that didn’t work.
I had some tricks in my user study, but ultimately, at least 25% of the data was completely ruined. This experience taught me quite a bit on data quality, checking data before you launch a large collection, and data cleaning.
I should have done better at checking data. My initial data collection showed some unexpected results:
- People open their mouth slightly when the start to walk. It is unintentional, but I definitely didn’t consider that people might want to breath while walking…
- No matter how slow you ask people to walk, they will mostly ignore you. At my frame rate, I needed people to walk less than 1/4 full walking pace (3 mph) to get enough frames, and that was pushing it. I could have used a conveyer belt, but that defeats the purpose.
- I should have aimed for quality not quantity. I think this was difficult to see clearly because other biometric datasets were getting so large at that time.
- I didn’t calibrate my setup every time because calibration took a long time. Some data paid the price later especially after someone kicked my setup by accident, and I had to fix everything. I should have taken more care to focus on something like that. There was an iterative process I could do with a ball to check how well the two light screens were calibrated to one another, but I didn’t figure trick out until the end of my collection.
My sensor collected raw data, and I worked on a pipeline to make a 3D reconstructed face. I didn’t work on making an overlay to the data or fit a surface to it. Looking back, that would have been a nice finishing touch. Even upsampling as I did a few years later would have done wonders to the appearance of the data and probably the algorithm’s performance.
Not making final prototype
I made 5 prototype sensors during graduate school, and I wish I had made a sixth. The sixth would have provided some interesting challenges and also would have provided an overload of biometric information.
The 6th Prototype
I have no regret getting into 3D image processing, biometrics, or computer vision. I don’t regret fighting to get the patent from Notre Dame and then fighting the Patent Evaluator to get it approved even though I started at Apple one month later after it was granted, so I didn’t use it. I don’t regret the strange path I went on to end up where I am now.
Further readings of mine: