[month] [year]

K N Amruth Sagar

Keralapura Nagaraju Amruth Sagar supervised by Dr. Santosh Ravi Kiran received his Master of Science by Research in Computer Science and Engineering (CSE). Here’s a summary of her research work on Ads and Anomalies: Structuring the Known and Probing the Unknown:

This thesis can be viewed in two main parts. In the first part, we introduce a new multilingual, hierarchical ad dataset, MAdVerse, designed to go beyond traditional plain class labels, which are often too simplistic to capture the full scope of brand identity and product type in an advertising context. By utilizing hierarchical labels, we enable a classification structure that reflects the natural hierarchy in which brands and products can be organized—a structure that is both more intuitive and more informative. This approach not only enriches the dataset with additional contextual information but also allows for more analytical insights. The above work naturally leads to a broader, intriguing question: Given the wide variation in visual styles, colors, and compositions across images, what specific visual attributes most influence a model’s ability to detect unfamiliar inputs? This leads us to Out-of-Distribution (OOD) detection—a field focused on identifying inputs that are novel or significantly different from the data a model was trained on. More specifically, we examine which visual attributes of an image most influence the OOD detection. Additionally, we investigate whether the choice of model backbone plays a significant role in these OOD outcomes, and explores whether these methods reveal any consistent patterns in how they respond to these features. Understanding these factors could deepen our insights into the robustness of detection methods across a range of visual contexts. By uncovering these influencing factors, we can better anticipate how detection methods handle this variability in attributes. Finally, we evaluate the effectiveness of hierarchical classifiers trained on our structured ad dataset, MAdVerse, when they encounter new images. This will allow us to observe how well these classifiers operate not only within established brand and product categories but also in managing the challenge of novel advertisements.

January 2025