A Generic View Planning System Based on Formal Expression of Perception Tasks.

View planning (VP) is a technique that guides the adjustment of the sensor's postures in multi-view perception tasks. It converts the perception process into active perception, which improves the intelligence and reduces the resource consumption of the robot. We propose a generic VP system for multiple kinds of visual perception. The VP system is built on the basis of the formal description of the visual task, and the next best view is calculated by the system. When dealing with a given visual task, we can simply update its description as the input of the VP system, and obtain the defined best view in real time. Formal description of the perception task includes the task's status, the objects' prior information library, the visual representation status and the optimization goal. The task's status and the visual representation status are updated when data are received at a new view. If the task's status has not reached its goal, candidate views are sorted based on the updated visual representation status, and the next best view that can minimize the entropy of the model space is chosen as the output of the VP system. Experiments of view planning for 3D recognition and reconstruction tasks are conducted, and the result shows that our algorithm has good performance on different tasks.

Continuous Viewpoint Planning in Conjunction with Dynamic Exploration for Active Object Recognition.

Active object recognition (AOR) aims at collecting additional information to improve recognition performance by purposefully adjusting the viewpoint of an agent. How to determine the next best viewpoint of the agent, i.e., viewpoint planning (VP), is a research focus. Most existing VP methods perform viewpoint exploration in the discrete viewpoint space, which have to sample viewpoint space and may bring in significant quantization error. To address this challenge, a continuous VP approach for AOR based on reinforcement learning is proposed. Specifically, we use two separate neural networks to model the VP policy as a parameterized Gaussian distribution and resort the proximal policy optimization framework to learn the policy. Furthermore, an adaptive entropy regularization based dynamic exploration scheme is presented to automatically adjust the viewpoint exploration ability in the learning process. To the end, experimental results on the public dataset GERMS well demonstrate the superiority of our proposed VP method.

Dual-NMS: A Method for Autonomously Removing False Detection Boxes from Aerial Image Object Detection Results.

In the field of aerial image object detection based on deep learning, it's difficult to extract features because the images are obtained from a top-down perspective. Therefore, there are numerous false detection boxes. The existing post-processing methods mainly remove overlapped detection boxes, but it's hard to eliminate false detection boxes. The proposed dual non-maximum suppression (dual-NMS) combines the density of detection boxes that are generated for each detected object with the corresponding classification confidence to autonomously remove the false detection boxes. With the dual-NMS as a post-processing method, the precision is greatly improved under the premise of keeping recall unchanged. In vehicle detection in aerial imagery (VEDAI) and dataset for object detection in aerial images (DOTA) datasets, the removal rate of false detection boxes is over 50%. Additionally, according to the characteristics of aerial images, the correlation calculation layer for feature channel separation and the dilated convolution guidance structure are proposed to enhance the feature extraction ability of the network, and these structures constitute the correlation network (CorrNet). Compared with you only look once (YOLOv3), the mean average precision (mAP) of the CorrNet for DOTA increased by 9.78%. Commingled with dual-NMS, the detection effect in aerial images is significantly improved.

Viewpoint planning with transition management for active object recognition.

Active object recognition (AOR) provides a paradigm where an agent can capture additional evidence by purposefully changing its viewpoint to improve the quality of recognition. One of the most concerned problems in AOR is viewpoint planning (VP) which refers to developing a policy to determine the next viewpoints of the agent. A research trend is to solve the VP problem with reinforcement learning, namely to use the viewpoint transitions explored by the agent to train the VP policy. However, most research discards the trained transitions, which may lead to an inefficient use of the explored transitions. To solve this challenge, we present a novel VP method with transition management based on reinforcement learning, which can reuse the explored viewpoint transitions. To be specific, a learning framework of the VP policy is first established via the deterministic policy gradient theory, which provides an opportunity to reuse the explored transitions. Then, we design a scheme of viewpoint transition management that can store the explored transitions and decide which transitions are used for the policy learning. Finally, within the framework, we develop an algorithm based on twin delayed deep deterministic policy gradient and the designed scheme to train the VP policy. Experiments on the public and challenging dataset GERMS show the effectiveness of our method in comparison with several competing approaches.