Learning naturalistic driving environment with statistical realism.

For simulation to be an effective tool for the development and testing of autonomous vehicles, the simulator must be able to produce realistic safety-critical scenarios with distribution-level accuracy. However, due to the high dimensionality of real-world driving environments and the rarity of long-tail safety-critical events, how to achieve statistical realism in simulation is a long-standing problem. In this paper, we develop NeuralNDE, a deep learning-based framework to learn multi-agent interaction behavior from vehicle trajectory data, and propose a conflict critic model and a safety mapping network to refine the generation process of safety-critical events, following real-world occurring frequencies and patterns. The results show that NeuralNDE can achieve both accurate safety-critical driving statistics (e.g., crash rate/type/severity and near-miss statistics, etc.) and normal driving statistics (e.g., vehicle speed/distance/yielding behavior distributions, etc.), as demonstrated in the simulation of urban driving environments. To the best of our knowledge, this is the first time that a simulation model can reproduce the real-world driving environment with statistical realism, particularly for safety-critical situations.

Dense reinforcement learning for safety validation of autonomous vehicles.

One critical bottleneck that impedes the development and deployment of autonomous vehicles is the prohibitively high economic and time costs required to validate their safety in a naturalistic driving environment, owing to the rarity of safety-critical events1. Here we report the development of an intelligent testing environment, where artificial-intelligence-based background agents are trained to validate the safety performances of autonomous vehicles in an accelerated mode, without loss of unbiasedness. From naturalistic driving data, the background agents learn what adversarial manoeuvre to execute through a dense deep-reinforcement-learning (D2RL) approach, in which Markov decision processes are edited by removing non-safety-critical states and reconnecting critical ones so that the information in the training data is densified. D2RL enables neural networks to learn from densified information with safety-critical events and achieves tasks that are intractable for traditional deep-reinforcement-learning approaches. We demonstrate the effectiveness of our approach by testing a highly automated vehicle in both highway and urban test tracks with an augmented-reality environment, combining simulated background vehicles with physical road infrastructure and a real autonomous test vehicle. Our results show that the D2RL-trained agents can accelerate the evaluation process by multiple orders of magnitude (103 to 105 times faster). In addition, D2RL will enable accelerated testing and training with other safety-critical autonomous systems.

Castle in the Sky: Dynamic Sky Replacement and Harmonization in Videos.

We propose a vision-based framework for dynamic sky replacement and harmonization in videos. Different from previous sky editing methods that either focus on static photos or require real-time pose signal from the camera's inertial measurement units, our method is purely vision-based, without any requirements on the capturing devices, and can be well applied to either online or offline processing scenarios. Our method runs in real-time and is free of manual interactions. We decompose the video sky replacement into several proxy tasks, including motion estimation, sky matting, and image blending. We derive the motion equation of an object at infinity on the image plane under the camera's motion, and propose "flow propagation", a novel method for robust motion estimation. We also propose a coarse-to-fine sky matting network to predict accurate sky matte and design image blending to improve the harmonization. Experiments are conducted on videos diversely captured in the wild and show high fidelity and good generalization capability of our framework in both visual quality and lighting/motion dynamics. We also introduce a new method for content-aware image augmentation and proved that this method is beneficial to visual perception in autonomous driving scenarios. Our code and animated results are available at https://github.com/jiupinjia/SkyAR.

Neural Rendering for Game Character Auto-Creation.

Many role-playing games feature character creation systems where players are allowed to edit the facial appearance of their in-game characters. This paper proposes a novel method to automatically create game characters based on a single face photo. We frame this "artistic creation" process under a self-supervised learning paradigm by leveraging the differentiable neural rendering. Considering the rendering process of a typical game engine is not differentiable, an "imitator" network is introduced to imitate the behavior of the engine so that the in-game characters can be smoothly optimized by gradient descent in an end-to-end fashion. Different from previous monocular 3D face reconstruction which focuses on generating 3D mesh-grid and ignores user interaction, our method produces fine-grained facial parameters with a clear physical significance where users can optionally fine-tune their auto-created characters by manually adjusting those parameters. Experiments on multiple large-scale face datasets show that our method can generate highly robust and vivid game characters. Our method has been applied to two games and has now provided over 10 million times of online services.

Adversarial Training for Solving Inverse Problems in Image Processing.

Inverse problems are a group of important mathematical problems that aim at estimating source data x and operation parameters z from inadequate observations y . In the image processing field, most recent deep learning-based methods simply deal with such problems under a pixel-wise regression framework (from y to x ) while ignoring the physics behind. In this paper, we re-examine these problems under a different viewpoint and propose a novel framework for solving certain types of inverse problems in image processing. Instead of predicting x directly from y , we train a deep neural network to estimate the degradation parameters z under an adversarial training paradigm. We show that if the degradation behind satisfies some certain assumptions, the solution to the problem can be improved by introducing additional adversarial constraints to the parameter space and the training may not even require pair-wise supervision. In our experiment, we apply our method to a variety of real-world problems, including image denoising, image deraining, image shadow removal, non-uniform illumination correction, and underdetermined blind source separation of images or speech signals. The results on multiple tasks demonstrate the effectiveness of our method.

Random Access Memories: A New Paradigm for Target Detection in High Resolution Aerial Remote Sensing Images.

We propose a new paradigm for target detection in high resolution aerial remote sensing images under small target priors. Previous remote sensing target detection methods frame the detection as learning of detection model + inference of class-label and bounding-box coordinates. Instead, we formulate it from a Bayesian view that at inference stage, the detection model is adaptively updated to maximize its posterior that is determined by both training and observation. We call this paradigm "random access memories (RAM)." In this paradigm, "Memories" can be interpreted as any model distribution learned from training data and "random access" means accessing memories and randomly adjusting the model at detection phase to obtain better adaptivity to any unseen distribution of test data. By leveraging some latest detection techniques e.g., deep Convolutional Neural Networks and multi-scale anchors, experimental results on a public remote sensing target detection data set show our method outperforms several other state of the art methods. We also introduce a new data set "LEarning, VIsion and Remote sensing laboratory (LEVIR)", which is one order of magnitude larger than other data sets of this field. LEVIR consists of a large set of Google Earth images, with over 22 k images and 10 k independently labeled targets. RAM gives noticeable upgrade of accuracy (an mean average precision improvement of 1% ~ 4%) of our baseline detectors with acceptable computational overhead.