Freshmen Research Immersion

Honors Student / Freshmen Researcher

Binghamton University - Binghamton, NY

~~~ Quick Description ~~~

  • Spent three semesters learning how to perform research in the honors program "FRI."

  • Developed a Deep Convolutional Inverse Graphics Network to segment litter from images.

  • Trained the DC-IGN in a semi-supervised manner.

  • Presented "Find the Litter: a Smart Way for Automatic Litter Detection" at a poster session.

~~~ In-Depth Description ~~~

        We trained the DC-IGN with 100 labeled images, then had it label our remaining 900 images. We then approved/disproved those 900 images. After approving, let's say 110 of those 900 images, we trained again - this time with 210 labeled images (the original 100 + the new 110). We then ran the new-and-improved network on the remaining unlabeled 690 images, and approved/disproved the results again. This led to a reinforcement-style learning loop, which was semi-supervised and allowed us to save time over labeling the data directly.

        Our semi-supervised learning method for the DC-IGN resulted in a segmentation algorithm that was superior to our best attempt at a computer vision solution for the same problem.

        The dataset we used was our own set of 1,000 pictures of litter around Binghamton's campus. The purpose of the project, aside from testing the semi-supervised learning idea on a DC-IGN, was to bring attention to this often-overlooked environmental problem at Binghamton University. I came up with the idea for the project after seeing trash on the ground all around the Union.

Screen Shot 2021-07-12 at 9.37.49 PM.png
Screen Shot 2021-07-12 at 9.37.49 PM.png
IMG_5259.HEIC