Mariana Sanz Planchart
Pronouns: She/her/hers
Research Mentor(s): Carol Flannagan
Co-Presenter: Luan, Hillary
Research Mentor School/College/Department: UMTRI / Engineering
Presentation Date: April 20
Presentation Type: Oral5
Session: Session 1 – 10am – 10:50am
Room: Breakout Room 1
Authors: Mariana Sanz Planchart, Carol Flannagan
Presenter: 4
Abstract
Despite the vast progress in the development of computer vision algorithms, algorithm bias is a growing concern among many researchers. Oftentimes, algorithm bias creeps in through the developmental phase as algorithms may be implemented using biased assumptions or they may be trained with skewed data sets that do not accurately simulate the real-world conditions in which they will be applied. This study attempts to mitigate algorithm bias by focusing on the improvement of the computer vision algorithm training sets used in a subset of transportation research projects. Labeling rubrics were developed to ensure inter-rater agreement and the consistent evaluation of face and cabin view videos that recorded several transportation behaviors such as people driving while distracted and passengers experiencing motion sickness. Videos of this type were used in an attempt to teach the algorithm to a) identify distracting, potentially dangerous behaviors people do while driving, and b) identify key markers of motion sickness in passengers engaging in a task. As a result of these procedures, the training data sets were diversified in size and scope. This is important because a less biased algorithm will perform better in the real world, benefiting people from underrepresented groups that, in the past, have been disproportionately affected by algorithm bias.
Engineering, Social Sciences
