Investigation of Factorized Optical Flows as Mid-Level Representations

TL;DR

A configurable framework is developed, along with four different environments that contain both static and dynamic objects, for analyzing the impacts of factorized optical flow maps on the performance of deep reinforcement learning agents.

Abstract

In this paper, we introduce a new concept of incorporating factorized flow maps as mid-level representations, for bridging the perception and the control modules in modular learning based robotic frameworks. To investigate the advantages of factorized flow maps and examine their interplay with the other types of mid-level representations, we further develop a configurable framework, along with four different environments that contain both static and dynamic objects, for analyzing the impacts of factorized optical flow maps on the performance of deep reinforcement learning agents. Based on this framework, we report our experimental results on various scenarios, and offer a set of analyses to justify our hypothesis. Finally, we validate flow factorization in real world scenarios.

Figures (7)

• Figure 1: A visualization of flow factorization. A raw flow map ($\mathcal{F}_{raw}$) can be factorized into an ego flow map ($\mathcal{F}_{ego}$) and an object flow map ($\mathcal{F}_{obj}$). The colors presented in the flow maps are drawn based on the directional color-coding as shown at the top-right corner. The vector at each pixel coordinate corresponds to a certain flow direction and magnitude.
• Figure 2: An overview of our framework, which allows incorporation and modification of environments, mid-level representations, scenarios, as well as DRL algorithms. Based on the given configuration, a set of mid-level representations are allowed to be selected from $\{\mathcal{S}, \mathcal{D}, \mathcal{F}_{raw}, \mathcal{F}_{ego}, \mathcal{F}_{obj}\}$, and form the state space $\mathbb{S}$ for the agent to learn a $\pi$ in a designated environment.
• Figure 3: An overview of the environments designed for our experiments. The green and red lines represent the walking paths of the pedestrians. The corresponding green and red suits denote the starting and ending zones of different paths, respectively.
• Figure 4: An impact analysis of the composition of the flow maps. The agents are trained using the normal speed mode ($1.2\sim1.8$ m/s), and evaluated under the low speed mode ($0.6\sim1.0$ m/s) and the high speed mode ($2.0\sim2.4$ m/s).
• Figure 5: The experimental results for validating the complementary property of the factorized flow maps (i.e., $\mathcal{F}_{ego}+\mathcal{F}_{obj}$).