With the development of Deep Reinforcement Learning (DRL), the applications of intelligent decisions in nonsymmetric information games have has become realizable. However, it is still a challenging task in DRL for its difficulties of building efficient exploration and action-reward mechanism, especially in an environment with multiple targets. To address this problem, a generating method of multi-target attacking strategy based on environment-aware DRL is proposed in this paper. Our proposed method consists of two stages in the agent learning process. The first stage is an environmentaware module for predicting the motion trajectories of multiple targets by using the optical flow estimation. The second stage is a decision-making module for predicting appropriate actions such as choosing angles and attacking by using the improved Deep Q-Network (DQN). To solving the problem of sparse rewards in the learning process, the motion trajectories predicted in the first stage are used to build reward trajectories for accelerating the convergence rate of the algorithm in the second stage. The experiments indicate that the proposed method can effectively generate multi-target attacking strategy in our self-built simulation environment. Our method can also provide a novel perspective of intelligent decisions in three-dimension space.
Autostereoscopic three-dimensional (3D) display has attracted considerable attention in recent years. To achieve high-quality 3D display with the lenticular-lens array, the accurate 3D display parameters of the lenticular-lens array need to be measured, including lines per inch (LPI) and slanted angle. Traditional methods generally measured the 3D display parameters by manually observing white-black characteristic 3D images synthesized based on the LPI and slanted angle. However, these methods have problems with time-consuming, inaccurate estimations, or limited applications. To address these problems, a method based on deep reinforcement learning (DRL) is proposed, which can automatically measure the 3D display parameters of the lenticular-lens array. In our method, the white-black characteristic 3D image is initialized based on the coarse LPI and slanted angle. Then the 3D image is captured by a camera from a 3D display device and input into a convolutional neural network (CNN) based on DRL. Finally, the CNN is used to optimize the LPI and slanted angle, and the accurate 3D display parameters can be measured. Experimental results show that our method can efficiently measure accurate 3D display parameters with roughly initial parameter values. We hope our study will make a valuable contribution to the field of autostereoscopic 3D display.
Continuous depth maps are reconstructed based on depth estimation from light-field data of axially distributed image sensing (ADS). In the proposed method, the light field of ADS is introduced, and the light-field trajectory function is presented, which formulates the relationship between image point positions in different axially captured images and the depth of object point. The depth value of an object is achieved by searching the light-field trajectory through statistical regression, which leads to an accurate depth estimation by making use of the structure information among the densely sampled views in light-field data. Moreover, as regions with smooth texture and occlusions do not contain reliable depth information, an energy minimization-based depth map optimization is carried out from initial estimations to obtain continuous and robust depth maps. Experimental results demonstrate that the reconstructed depth maps are accurate, continuous, and able to preserve more details. Moreover, the synthesized light-field images calculated by reconstructed depth map show good effect in 3D light-field display, and the validity of the proposed method is verified.
KEYWORDS: 3D modeling, 3D displays, Reconstruction algorithms, Optical flow, New and emerging technologies, Evolutionary algorithms, Neural networks, Cameras, 3D image processing, Image segmentation
In recent years, 3D display technique is one of the emerging technology and gradually becomes accessible to a broader audience. However, because of the traditional 3D reconstruction method is limited by the number of the feature found in the image, the resolution of the generated 3D model is not high enough for 3D display. A new system is purposed, in which we consider the vertical and horizontal disparity between images, and the optical flow is used to replace the feature matching segment, so that more points can be pushed into the reconstruction process for improving the resolution of the models. Experimental results prove that the resolution of the models can be enhanced effectively. The details of the model are preserved, and the holes in the weak texture region are successfully filled.
The deep convolution neural network has been widely tackled for optical flow estimation in recent works. Due to advantages of extracting abstract features and efficiency, the accuracy of optical flow estimation using CNN is improved steadily. However, the edge information for most flow predictions is vague. Here, two methods are presented to add extra useful information in training our optical flow network, for the purpose of enhancing edge information of the result. The edges map is added into the input section, and the motion boundary is considered for the input section. Experimental result shows that the accuracy with both methods is higher than the control experiment. 3.71% and 7.54% are improved by comparing just a pair of frames in the input section respectively.
Light-field cameras are used in consumer and industrial applications. An array of micro-lenses captures enough information that one can refocus images after acquisition, as well as shift one’s viewpoint within the sub-apertures of the main lens, effectively obtaining multiple views. Thus, depth estimation from both defocus and correspondence are now available in a single capture. And Lytro.Inc also provides a depth estimation from a single-shot capture with light field camera, like Lytro Illum. This Lytro depth estimation containing many correct depth information can be used for higher quality estimation. In this paper, we present a novel simple and principled algorithm that computes dense depth estimation by combining defocus, correspondence and Lytro depth estimations. We analyze 2D epipolar image (EPI) to get defocus and correspondence depth maps. Defocus depth is obtained by computing the spatial gradient after angular integration and correspondence depth by computing the angular variance from EPIs. Lytro depth can be extracted from Lyrto Illum with software. We then show how to combine the three cues into a high quality depth map. Our method for depth estimation is suitable for computer vision applications such as matting, full control of depth-of-field, and surface reconstruction, as well as light filed display
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.