BeVIS Benchmark Dataset

1. BeVIS Benchmark Dataset

(pw: 2bn0)

Note that since the camera id assigned by the system might be different at each runtime, the direction information in the image name might not match the real direction. In the extraCalib/Surround view System folder, the image labeled Left is an image by the Back camera, the image labeled Back is a Right image, and the image labeled Right is a Left image.

As for IMU data, it contains const double& gx, const double& gy, const double& gz, const double& ax, const double& ay, const double& az, const double& t. And vector_t in data.cc is defined as: typedef std::vector<IMUData, Eigen::aligned_allocator<IMUData>> vector_t.

Currently in the BeVIS, there are totally 34,951 front-view images, 34,951 IMU motion data segments with each segment recording the vehicle motion between each two consecutive front-view images, and 12,032 surround-view images (synthesized from 48,128 fisheye images) collected in four indoor parking environments at Tongji University, covering a wide variety of real cases for autonomous indoor parking. One advantage of the BeVIS is the diversity of the conditions for data collection, ranging from static scenes with bright illumination to dynamic scenes with poor illumination. To exploit this advantage, we manually classify all the data in the BeVIS into three levels, “easy”, “medium” and “difficult”. In addition, another outstanding advantage of the BeVIS is the availability of the groundtruth trajectories for each sequence, ensuring objective evaluation of different SLAM systems for autonomous indoor parking.

2. The BeVIS is only for acamedic use under the GNU General Public License Version 3 (GPLv3). For commercial purposes, please contact the authors for details.