Shubham Trehan

PUBLICATIONS

Towards Active Vision for Action Localization with Reactive Control and Predictive Learning Shubham Trehan and Sathyanarayanan N. Aakur [WACV 2022]

In this groundbreaking research, we introduce an innovative approach to active action localization, where an active camera dynamically adjusts to keep actions within its field of view.

Key Highlights

• Active Camera Control: Our system dynamically adjusts the camera’s orientation to follow the action, reducing the chances of losing track due to occlusion or rapid movements.

• Predictive Learning: By anticipating future frames, our model maintains a continuous focus on the action, enhancing tracking accuracy.

• Real-World Applications: From surveillance to assistive robotics, our approach promises significant improvements in scenarios where maintaining the visual focus on moving subjects is critical.

ProtoKD: Learning from Extremely Scarce Data for Parasite Ova Recognition Shubham Trehan Udhav Ramachandran Ruth Scimeca Sathyanarayanan N. Aakur [ICMLA 2023]

Key Highlights

• Prototypical Networks: Leverages the strength of Prototypical Networks coupled with Knowledge distillation to create robust class prototypes, facilitating effective few-shot learning.

• Application to Medical Datasets: Specifically designed to handle scenarios with extremely limited data, demonstrating significant improvements in medical image classification tasks.

Discovering Novel Actions from Open World Egocentric Videos with Object-Grounded Visual Commonsense Reasoning Sanjoy Kundu Shubham Trehan and Sathyanarayanan N. Aakur [ECCV 2024]

Dive into our innovative research on discovering novel actions from egocentric videos using a neuro-symbolic framework. This work combines the power of object-centric vision-language models and large-scale knowledge bases to achieve unparalleled performance in open-world activity inference.

Key Highlights

• Neuro-Symbolic Framework: Integrates symbolic knowledge and visual data to ground objects and infer actions in open-world settings.

• Object-Driven Activity Discovery: Uses prior knowledge to discover plausible actions, leveraging object affordances for accurate activity recognition.

Self-supervised Multi-actor Social Activity Understanding in Streaming Videos Shubham Trehan and Sathyanarayanan N. Aakur [ICPR 2024]

Explore our latest research on social activity recognition that tackles the complexities of detecting and understanding simultaneous actions and social interactions among multiple actors in streaming videos. This self-supervised framework leverages multi-actor predictive learning and visual-semantic graphs to deliver robust performance with minimal labeled data.

Key Highlights

• Self-Supervised Learning: Achieves high accuracy in social activity recognition without relying on densely annotated data.

• Multi-Actor Predictive Learning: Jointly models individual actions and group-level activities, enhancing the understanding of social interactions.

• Visual-Semantic Graphs: Uses graph structures to represent social interactions, enabling relational reasoning and improved performance on standard benchmarks.