Shulou(Shulou.com)06/02 Report--

Author: Christopher Dossman compilation: Jieqiong, Conrad, Yun Zhou Hello, everyone. This week's AI Scholar Weekly column meets you again. AI Scholar Weekly is an academic column in the field of AI, dedicated to bringing you the latest, most comprehensive and in-depth academic overview of AI, capturing all the cutting-edge information of AI learning every week. At the end of the article, we will update the AI black mirror series of short stories every week to do AI research. This is enough from the beginning of this week's keywords: ACRV robot vision challenge, vision-based mobile operation, Sim2Real strategy, DC-SPP-YOLO model

Hot academic research this week

New probabilistic target detection challenge: the first ACRV robot vision challenge.

For robot applications, answering the question of what the target is and where the target is, and providing a measure of spatial and semantic uncertainty, is a priority problem to be solved in target detection.

A few days ago, the Robot Vision Excellence Center of the Australian Research Council, supported by Google, opened the first challenge on probabilistic target detection. The challenge of computer and robot vision requires participants to detect objects in video data and provide accurate estimates of spatial and semantic uncertainty.

Figure 1: the sample picture is from the simulation environment used to generate challenge test data. The first row and the bottom left are the environment for the test sequence, and the bottom right is the verification sequence.

There is no threshold for this challenge, and anyone in the artificial intelligence community who is interested in target detection can participate, which is a very good challenge. The test data set for the challenge, which contains more than 56000 images from 18 simulated indoor video sequences, will be evaluated on a public server for the challenge, which is open only during the open competition phase. Participants will be ranked and share a bonus of A $5000.

Potential application and effect

This new challenge is the introduction of probabilistic target detection, which promotes the existing target detection tasks to the spatial and semantic uncertainty in high-end robot applications. Generally speaking, it will improve the technical level of robot application in object detection.

More details:

Https://nikosuenderhauf.github.io/roboticvisionchallenges/object-detection

Original text:

Https://arxiv.org/abs/1903.07840

A Vision-based Image Generation method for Mobile operating Robot

The researchers designed a system that uses a single input image to generate a set of images for a specific object from a desired point of view for use by a mobile robot.

The proposed method is a deep neural network that trains it to "imagine" the appearance of an object from multiple perspectives. It takes a single RGB image of the object as input and returns a set of RGB and depth images (bit depth images), eliminating traditional, time-consuming scanning.

Although the deep neural network has achieved single-view reconstruction, it is difficult to obtain the accurate details of the reconstruction object directly because of the large amount of computation in the reconstruction process. In this method, a target detector based on CNN model is used to extract the target from the natural environment, and a set of RGB and depth images are generated by the neural network (bit depth image, the current bit depth in Ps software is 16 bits). This method has been tested on the generated image and the real image, and it is proved to be very effective.

Potential application and effect

Image-based generation has the potential to provide better spatial resolution for reconstructed objects. Therefore, this method is necessary in the field of mobile operating robot. This method may help the robot better understand the spatial properties of an object without the need for a complete scan.

Original text:

Https://arxiv.org/abs/1903.06814

Enhance the composite image for Sim2Real strategy migration

Because there is a domain gap between the actual data and the synthetic data, and it is difficult to transfer the strategies learned in the simulator to the real scene. In the past, domain randomization (random domain data generation, domain randomization) solved this challenge by using random transformations such as random object shapes and textures to enhance synthetic data.

In recent years, researchers have made new contributions to the research of domain randomization, optimizing the enhancement technology of Sim2Real migration, so that it can achieve domain-independent strategy learning without real images.

They designed an efficient search method for depth image enhancement using target location. In the process of strategy learning, the generated random transformation sequence is used to enhance the synthetic depth image.

In order to assess the extent of this migration, the researchers proposed a delegated task for target location estimation, which requires very little real data. The new method greatly improves the accuracy of evaluating operation tasks on real robots.

Potential application and effect

This method promotes the effective learning of operation strategy in the simulation environment. This is very useful because the simulator promotes scalability and provides access to the underlying space during model training. In addition, the new method does not need real images to achieve strategy learning and can be applied to a variety of operational tasks.

Original text:

Https://arxiv.org/abs/1903.07740

The new DC-SPP-YOLO model realizes more accurate real-time target detection.

The researchers proposed a DC-SPP-YOLO method to improve the accuracy of YOLOv2 target detection. The DC-SPP method improves YOLOv2 by optimizing the connection structure of the basic network, and introduces multi-scale local area feature extraction. Therefore, the proposed new method is more accurate than YOLOv2.

It achieves the target detection speed close to YOLOv2 and is higher than traditional target detection methods such as deconvolution single shot detector (DSSD), scale transferable detection network (STDN) and YOLOv3.

DC-SPP-YOLO especially uses the connection of the convolution layer in the YOLOv2 basic network to enhance feature extraction and minimize the vanishing gradient problem. On this basis, an improved spatial pyramid pool is proposed, and the multi-scale local region features are connected together, so that the network can learn the target features more comprehensively.

Based on a new loss function training DC-SPP-YOLO model, the loss function is composed of mean square error and cross entropy, which can realize target detection more accurately. The experimental results show that the mAP of DC-SPP-YOLO on PASCAL VOC and UA-DETRAC datasets is larger than that of YOLOv2.

Potential use and impact

By strengthening feature extraction and using multi-scale local features, DC-SPP-YOLO achieves better real-time target detection accuracy than YOLOv2. This method can be used to realize more accurate and advanced computer vision applications in safety monitoring, medical diagnosis, autopilot and so on.

For details, please refer to:

Https://arxiv.org/abs/1903.08589

Interactive Medical Image Segmentation based on full convolution Neural Network

Recent research has proposed an "intelligent" semi-automatic segmentation method for deep learning, which can interactively describe the regions of interest in medical images. The proposed method uses a FCNN architecture to perform interactive two-dimensional medical image segmentation.

So how does it use interaction? The network is trained to divide only one area of interest at a time, taking into account what the user enters in the form of one or more mouse clicks. The model is also trained to use the original 2D image and a "pilot signal" as input. It then outputs the binary mask of the specific split object. Researchers have shown that it can be used to segment various organs in abdominal CT. This new method provides very accurate results and can be corrected in a fast, intelligent and spontaneous manner according to the user's choice.

Potential use and impact

This method can quickly provide high-end two-dimensional segmentation results. It also has the potential to solve urgent clinical challenges and can be used in many medical imaging applications to improve the accuracy of segmentation, such as tumor localization, surgical planning, diagnosis, intraoperative navigation, virtual surgical simulation, tissue volume measurement and so on. Other applications include visualization, radiotherapy planning, 3D printing, image classification, natural language processing, and so on.

Original text:

Https://arxiv.org/abs/1903.08205

Other popular style papers

3D point clouds are used to enhance the environment classification of wearable robots.

Original text:

Https://arxiv.org/abs/1903.06846v1

A pre-training model for real-time semantic segmentation of road driving images.

Original text:

Https://arxiv.org/abs/1903.08469

The first event-based motion segmentation data set learning method and the learning pipeline of the event camera.

Original text:

Https://arxiv.org/abs/1903.07520

New research shows that you can improve the accuracy of visual SLAM algorithms through Wi-Fi sensing. Original: https://arxiv.org/abs/1903.06687

Plug and play magnetic resonance imaging (MRI).

Original text:

Https://arxiv.org/abs/1903.08616

AI News

Google posted AI graffiti in memory of German composer Bach on his birthday.

For more information, please see First ever AI doodle that allows users to make music.

Https://www.newsweek.com/google-doodle-bach-birthday-when-march-21-22-1366826

Combine compute-intensive artificial intelligence applications with recently released new artificial intelligence servers.

For more information, please see: AI Server Enabled with NVIDIA GPUs for edge computing

Https://www.marketwatch.com/press-release/inspur-releases-edge-computing-ai-server-enabled-with-nvidia-gpus-2019-03-19?mod=mw_quote_news

Is it really possible to create artificial intelligence similar to human beings?

For more information, please see: How to create AI that is more human

Introduction by https://www.forbes.com/sites/jenniferhicks/2019/03/19/how-do-we-create-artificial-intelligence-that-is-more-human/amp/ columnist

Christopher Dossman is the chief data scientist at Wonder Technologies and has lived in Beijing for five years. He is an expert in the deployment of deep learning systems and has extensive experience in developing new AI products. In addition to his excellent engineering experience, he also taught 1000 students the basics of deep learning.

LinkedIn:

Https://www.linkedin.com/in/christopherdossman/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.