Visual Recognition and Synthesis of Human-Object Interactions

Abstract

The ability to perceive and understand people's actions enables humans to efficiently communicate and collaborate in society. Endowing machines with such ability is an important step for building assistive and socially-aware robots. Despite such significance, the problem poses a great challenge and the current state of the art is still nowhere close to human-level performance. This dissertation drives progress on visual action understanding in the scope of human-object interactions (HOI), a major branch of human actions that dominates our everyday life. Specifically, we address the challenges of two important tasks: visual recognition and visual synthesis.

The first part of this dissertation considers the recognition task. The main bottleneck of current research is a lack of proper benchmark, since existing action datasets contain only a small number of categories with limited diversity. To this end, we set out to construct a large-scale benchmark for HOI recognition. We first tackle the problem of establishing the vocabulary for human-object interactions, by investigating a variety of automatic approaches as well as a crowdsourcing approach that collects human labeled categories. Given the vocabulary, we then construct a large-scale image dataset of human-object interactions by annotating web images through online crowdsourcing. The new "HICO" dataset surpasses prior datasets in term of both the number of images and action categories by one order of magnitude. The introduction of HICO enables us to benchmark state-of-the-art recognition approaches and also shed light on new challenges in the realm of large-scale HOI recognition. We further discover that visual features of humans, objects, as well as their spatial relations play a central role in the representation of interaction, and the combination of three can improve the recognition outcome.

The second part of this dissertation considers the synthesis task, and focuses particularly on the synthesis of body motion. The central goal is: given an image of a scene, synthesize the course of an action conditioned on the observed scene. Such capability can predict possible actions afforded by the scene, and will facilitate efficient reactions in human-robot interactions. We investigate two types of synthesis tasks: semantic-driven synthesis and goal-driven synthesis. For semantic-driven synthesis, we study the forecasting of human dynamics from a static image. We propose a novel deep neural network architecture that extracts semantic information from the image and use it to predict future body movement. For goal-directed synthesis, we study the synthesis of motion defined by human-object interactions. We focus on one particular class of interactions—a person sitting onto a chair. To ensure realistic motion from physical interactions, we leverage a physics simulated environment that contains a humanoid and chair model. We propose a novel reinforcement learning framework, and show that the synthesized motion can generalize to different initial human-chair configurations.

At the end of this dissertation, we also contribute a new approach to temporal action localization, an essential task in video action understanding. We address the shortcomings of prior Faster R-CNN based approaches, and show state-of-the-art performance on standard benchmarks.