Document Type

Presentation

Publication Date

3-1-2024

Publication Title

Chicago Society for Neuroscience Annual Meeting

Publisher Name

Chicago Society for Neuroscience

Publisher Location

Chicago

Abstract

Object recognition is a crucial function of biological vision; it allows us to draw conclusions about a visual scene that transcends the image formed by the retina. However, the task of object recognition quickly becomes a challenge when hindrances such as viewing angle, object distance from observer, illuminant qualities, and potential occlusions become active variables. Additionally, the diversity of visual features within the same category of object, coupled with the numerous contexts in which an object may be observed is demonstrative of the formidable task that is object recognition. Previous research showed a significant texture bias in Convolutional Neural Networks’ object recognition while humans recognize shape despite textural differences. We first sought to create a dataset composed of novel stimuli comprised of 100 3-dimensional models, with 50 biological models and 50 non-biological models, with 10 models in each of the 10 object categories. We then collected 100 texture images that corresponded with the models (i.e.,10 butterfly models and 10 butterfly textures). These models and images were then cross-linked via script to generate every possible combination of these texture wrapped models. This algorithm generated 10,000 unique models wherein both nondiagnostic and diagnostic textures were mapped. This same algorithm then proceeded to render an image every 30 degrees along the z-axis for a total of 12 different viewpoints of each model. This results in a dataset, deemed ReTexture, consisting of 120,000 images. The ReTexture dataset was then utilized to create a human experiment, wherein a human’s ability to recognize these models with varying textures at different viewpoints was evaluated. The human experiment of our study is essential, as we must have a baseline for human object recognition to compare these findings to Convolutional Neural Networks’ object recognition on the same models. Following the human experiment, the ReTexture dataset was then run through a variety of neural networks for object recognition and the data was sieved through to compare whether the networks gave greater biases to object shape or object texture.

This study is funded by the Carbon Research Fellowship at Loyola University Chicago awarded to Luke D. Baumel for 2022-2024

Comments

Author Posting © The Authors, 2024. This poster was presented at the Chicago Society for Neuroscience Annual Meeting on March 1st, 2024. https://chicagosfn.org/annual-meeting/chicago-chapter-2024-annual-meeting/

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.

Download

Share

COinS