The Challenge of Appearance-Free Object Tracking with Feedforward Neural Networks

• DOI: 10.48550/arxiv.2110.02772
• Publication Date: 2021-01-01
• Authors: Girik Malik; Drew Linsley; T. Serre; Ennio Mingolla
• Keywords: Feed forward; Tracking (education)

Nearly all models for object tracking with artificial neural networks depend on appearance features extracted from a "backbone" architecture, designed recognition. Indeed, significant progress has been spurred by introducing backbones that are better able to discriminate objects their appearance. However, extensive neurophysiology and psychophysics evidence suggests biological visual systems track using both motion features. Here, we introduce $\textit{PathTracker}$, challenge inspired cognitive psychology, which tests the ability of observers learn solely motion. We find standard 3D-convolutional deep network struggle solve this task when clutter is introduced into generated scenes, or travel long distances. This reveals tracing path blind spot feedforward networks. expect strategies appearance-free vision can inspire solutions these failures