Tanner Watts
I am a Machine Learning and Robotics Research Scientist at the University of Utah, where I work on creating autonomous surgical control algorithms.
PhD candidate in Robotics (Deep Learning), University of Utah
B.C.S Computer Science and B.S in Applied Mathematics, University of Utah
Research
Understanding the physical world is one of the most fascinating applications of Deep Learning. My research interests include applications of these powerful models in surgical scenes, biology, and agriculture.
Publications
DefGoalNet: Contextual Goal Learning from Demonstrations For Deformable Object Manipulation
Bao Thach*, Tanner Watts*, Shing-Hei Ho, Tucker Hermans, Alan Kuntz
* = Co-First Author
DefGoalNet generates a goal point cloud based on both simulated and human demonstration. This model combined with deformerNet is able to fully autonomously complete multiple surgical tasks.
CountNet3D: A 3D Computer Vision Approach to Infer Counts of Occluded Objects
Porter Jenkins, Kyle Armstrong, Stephen Nelson, Siddhesh Gotad, J. Stockton Jenkins, Wade Wilkey, Tanner Watts
Published in WACV 2023. IEEE
My main contrubtion to this paper was an algorithim that can map the 2D vision bottle labels to the 3D cans contained in the point beams.
Projects
OLSeg
I created this project to help neurobiology labs during my time at Stanford University (2022). It utilizes the MaskRCNN from Meta, for instance segmentation. The major challenge of this project was creating the most accurate possible masks for these cells.
There is amazing information on the shape and size of oligodendrocytes, but precision is required. A form of pixel-wise BFS was used to make these masks, this is evident in the code.
Serotonin Segmentation
This project was also created to help neurobiology labs during my time at Stanford University (2022). It utilizes the MaskRCNN from Meta, for instance segmentation. Unlike OLSeg the precision of the border is less important, but the count is very important. This project uses dectron2 to count fos cells.
Line Tracking
Understanding the wait times of individuals waiting to order food is an interesting machine-learning problem. While at Enzy, I used a segmentation model similar to the two above, adding a Kalman box tracking algorithm to time individuals as they move across the screen successfully.
Scaling this algorithm was a major challenge, I learned how to use docker, and AWS to port my algorithm to the cloud.
Football Analysis in Real Time using Transformer
With transformer models becoming more and more popular, in the fall of 2022 I decided to implement a visual transformer with the goal to create a model that can correctly identify key moments in soccer games. While the transformer works, the lack of data was a major challenge that was difficult to overcome. Even though the model does not fit the data, it was interesting to see the performance of ResNet vs a visual transformer.
Extra Curricular
Ski Instructor with PCSS (Park City Ski and Snowboard)
Bolt Ranch and Bolt Ranch Store
- Created Logos and websites for the store.
- Continue to work with cattle (herding, vacinating, branding and maintining)
