Research Conducted during my internship at SDSC (San Diego Super Computer Center)
Working under Dr. Paul Rodriguez, we were given 171,000 Depression-Era Photos. We were supposed to extract valuable information from these images. The issue was, we had far too many images, and even with access to a supercomputer, it would take forever to test out different methods on the dataset. This meant we needed to test our algorithms first locally, and later, test it on the larger set.
- Object Recognition
- Face Detection
Looking at the dataset, we quickly realized that each image had a border around them, which we would need to remove.
After looking at the images, I realized that there were straight lines at the top and bottom of the images right before the border. Thus, I employed the usage of the Hough Lines transform in order to find all straight lines. From there, I figured out which were the top/bottom most lines and which were the left/right most lines. Then, I cropped.
After first looking for different object recognition softwares, I finally settled on Darknet.
Darknet, more specifically, the YOLO algorithm, allowed us to quickly and easily perform object recognition. By making "predictions with a single network evaluation", Darknet's YOLO is able to perform extremely fast (as it relies on only one NN per image). (Redmon)
We deployed the program on a small sample size of 90 images, and found it having an 82% predictive value.
| True | False | |
|---|---|---|
| True | 74 (correctly found the object) | 13 (didn't find the object) |
| False | 3 (incorrectly indentified the object) |
Paul Rodriguez (Amazing Mentor) Ethan Yao (My partner, did the Face Detection)