Amazon Grocery Vision Challenge
1st Place Winner

Our award-winning solution for the ICCV 2025 Grocery Vision Challenge represents a breakthrough in understanding complex human behaviors in retail environments. We developed a sophisticated multi-modal fusion architecture that integrates cutting-edge deep learning techniques to achieve unprecedented accuracy in detecting and classifying three critical consumer behaviors: Take, Return, and Rummage actions.

The solution combines Temporal Action Localization (TAL) and Spatio-Temporal Action Localization (STAL) approaches, leveraging the OpenTAD framework with AdaTAD and CausalTAD models, alongside GLEE object detection and SAM2 segmentation for comprehensive scene understanding.