Virtual Worlds as Proxy for Multi-Object Tracking Analysis

TL;DR

This paper introduces a method to generate realistic virtual worlds for training and evaluating computer vision algorithms, demonstrating that virtual data can effectively substitute real data for object tracking and scene understanding tasks.

Contribution

It presents a novel real-to-virtual world cloning approach and releases a new labeled dataset, showing virtual data enhances deep learning performance and allows controlled testing of environmental factors.

Findings

Pre-trained models perform similarly in real and virtual worlds.

Pre-training on virtual data improves real-world performance.

Weather and imaging conditions significantly impact tracking accuracy.

Abstract

Modern computer vision algorithms typically require expensive data acquisition and accurate manual labeling. In this work, we instead leverage the recent progress in computer graphics to generate fully labeled, dynamic, and photo-realistic proxy virtual worlds. We propose an efficient real-to-virtual world cloning method, and validate our approach by building and publicly releasing a new video dataset, called Virtual KITTI (see http://www.xrce.xerox.com/Research-Development/Computer-Vision/Proxy-Virtual-Worlds), automatically labeled with accurate ground truth for object detection, tracking, scene and instance segmentation, depth, and optical flow. We provide quantitative experimental evidence suggesting that (i) modern deep learning algorithms pre-trained on real data behave similarly in real and virtual worlds, and (ii) pre-training on virtual data improves performance. As the gap between real and virtual worlds is small, virtual worlds enable measuring the impact of various weather and imaging conditions on recognition performance, all other things being equal. We show these factors may affect drastically otherwise high-performing deep models for tracking.