Active Observers in a 3D World: Human Visual Behaviours for Active Vision

Human-like performance in computational vision systems is yet to be achieved. In fact, human-like visuospatial behaviours are not well understood – a crucial capability for any robotic system whose role is to be a real assistant. This dissertation examines human visual behaviours involved in solving a well-known visual task; The Same-Different Task. It is used as a probe to explore the space of active human observation during visual problem-solving. It asks a simple question: “are two objects the same?”. To study this question, we created a set of novel objects with known complexity to push the boundaries of the human visual system. We wanted to examine these behaviours as opposed to the static, 2D, display-driven experiments done to date. We thus needed to develop a complete infrastructure for an experimental investigation using 3D objects and active, free, human observers. We have built a novel, psychophysical experimental setup that allows for precise and synchronized gaze and head-pose tracking to analyze subjects performing the task. To the best of our knowledge, no other system provides the same characteristics. We have collected detailed, first-of-its-kind data of humans performing a visuospatial task in hundreds of experiments. We present an in-depth analysis of different metrics of humans solving this task, who demonstrated up to 100% accuracy for specific settings and that no trial used less than six fixations. We provide a complexity analysis that reveals human performance in solving this task is about O(n), where n is the size of the object. Furthermore, we discovered that our subjects used many different visuospatial strategies and showed that they are deployed dynamically. Strikingly, no learning effect was observed that affected the accuracy. With this extensive and unique data set, we addressed its computational counterpart. We used reinforcement learning to learn the three-dimensional same-different task and discovered crucial limitations which only were overcome if the task was simplified to the point of trivialization. Lastly, we formalized a set of suggestions to inform the enhancement of existing machine learning methods based on our findings from the human experiments and multiple tests we performed with modern machine learning methods