Improving the Motion Processing Hierarchy for Attending to Visual Motion

Abstract

Visual motion has been studied for decades now. Attention to motion using Selective Tuning involves a top-down selection mechanism within a feed-forward motion hierarchy. Researchers have proposed various models for the motion hierarchy. In this thesis, we introduce a learnable hierarchy, based on fully convolutional networks, ST-Motion-Net. The Selective Tuning model for visual attention is demonstrated on ST-Motion-Net to localize motion patterns and segment moving objects. We create two datasets, Blender-MP and Blender-Complex, to evaluate ST-Motion-Net on motion pattern detection, localization, and motion segmentation tasks. ST-Motion-Net achieves excellent performance on motion pattern detection and localization for each area of ST-Motion-Net. For motion segmentation, we evaluate 2-Frame-Area-V1 of ST-Motion-Net on the task. 2-Frame-V1 contains neurons that respond to translation motion, given 2 most recent frames of a temporal sequence. 2-Frame-V1 achieves 86.84% IoU on Blender-MP-Test, which surpass some state-of-the-art models. On Blender-Complex-Test, 2-Frame-V1 reaches 52.61% IoU, which also achieves state-of-the-art performance.