3D target detection data set acquisition device and marking method

Abstract

The invention discloses a device for collecting a 3D target detection image data set and a labeling method, and relates to the technical field of 3D target detection data set devices. The device comprises a bottom rotating device, a horizontal telescopic device, a vertical lifting device, a pitching angle adjusting device, a camera clamping and rotating device and a control module. The control module comprises five motors, a main control board and a power supply module, each motor can drive different devices to move so as to realize all corresponding poses of the camera and a to-be-collected target on a hemispherical surface within a certain range, and the main control board can control the motors to make the motors move in a coordinated manner and make camera image collection; and the relative pose of the camera and the target can be obtained according to the state of each motor, so that the coordinate of the object under the camera can be obtained, and the object is marked in the picture. According to the method, the 3D target detection data set manufacturing and labeling process can be greatly simplified.

Description

technical field [0001] The present invention relates to an image data collection device and collection method, in particular to a 3D target detection image data collection device and collection method. Background technique [0002] 3D object detection and pose estimation are research hotspots in recent years, and play an irreplaceable role in the fields of robot intelligent grasping, automatic driving and augmented reality. With the rapid development of deep learning, it has become an irreversible trend to use deep learning technology to realize 3D object detection and pose estimation. However, deep learning technology is inseparable from a large amount of data and needs to be labeled. At this stage, data sets are divided into software virtual synthesis and real scene collection: synthetic data sets have poor generalization; most data sets in real scenes rely on industrial machinery. The arm needs to be labeled with Aruco codes, which has the disadvantages of complex proces...

AI Technical Summary

Problems solved by technology

However, deep learning technology is inseparable from a large amount of data and needs to be labeled. At this stage, data sets are divided into software virtual synthesis and real scene collection: synthetic data sets have poor generalization; most data sets in real scenes rely on industrial machinery. The arm needs to be labeled with Aruco codes, which has the disadvantages of complex process, high hardware requirements, cumbersome and inaccurate labeling, and a large number of Aruco codes occupying the image background.

The collection and labeling of 3D target detection and pose estimation image datasets in real scenes has brought great troubles to a large number of researchers, making it impossible for scholars to spend more time on 3D target detection and pose estimation algorithms. Unable to apply more researched algorithms to practice

Images