The Sadra Fort (SF) Dataset

The SF dataset has been recorded at the Sadra Fort archaeological site, Gujarat, India. Located on the banks of the river Sabarmati, the fort was built in the 15th Century and is spread across an area of more than 10000 sq. meters. The main site consists of a small palace which is covered with walls on three sides. The SF dataset consists of four sequences, each covering a different part of the fort. While the SF-2, SF-3, and SF-4 focus on mapping the three walls, the SF-1 is recorded for mapping the palace. The characteristics of the dataset are shown in the Table below. The length of the trajectories lie between 100 to 300 meters.

A DJI Mavic Pro has been used for preparing the SF monocular camera dataset. The Mavic Pro is a quadrotor with a 3-axis gimbal stabilization for the monocular camera. The dataset is recorded at a resolution of 1280x720 pixels at 30 Hz. The DJI Mavic Pro GPS has a vertical accuracy of ± 0.5 meters and a horizontal accuracy of ± 1.5 meters. Using the GPS and the roll, pitch, and yaw data from the flight record, we provide an approximate trajectory ground truth in a world coordinate frame of reference whose origin is fixed at the same latitude, longitude, and altitude for all the sequences. We provide the sequences in a ROS bag file containing the video sequence at 30fps and the approximate ground-truth pose at synchronized 10fps.

Sequence
Name
Duration (s)Trajectory
length (m)
Size (GB)Download Link
SF-137019816.2SF-1.bag
SF-232622315.1SF-2.bag
SF-339929617.1SF-3.bag
SF-42001178.67SF-4.bag

  • Download all sequences as a zip folder from here. [61.2GB]
  • Download the Groundtruth of the sequences from here. [522 KB]
  • Download the Flight Data of the sequences from here. [1.4 MB]
  • The groundtruth of the sequences has been provided in text files in a format compatible with the RPG Trajectory Evaluation.
  • The dataset has been recorded at the Sadra Fort, Gujarat, India at a resolution of 1280 x 720 (BGR) at 30 Hz.
  • Refer to our Github repository for the camera calibration file.
  • Use Rosbag Decompress to get the uncompressed ROS bag files using the following command-

$ rosbag decompress *.bag

  • Note: The Gimbal pitch has been changed mid-flight in some of the sequences and hence you may refer to the flight data to get that information. You may create a transformation using that data.

The details of the bag file topics have been provided below-

/camera/image_raw

Msg Type: sensor_msgs/Image, The video frames are published on this topic at 30Hz frame rate

/dji/gimal_pitch

Msg Type: geometry_msgs/Vector3Stamped, The gimbal pitch is published in degrees in the y component of the vector

/dji/gps_fix

Msg Type: sensor_msgs/NavSatFix, The latitude, longitude, and altitude of the UAV is published on this topic

/dji/groundtruth

Msg Type: geometry_msgs/PoseWithCovarianceStamped, The groundtruth of the sequence at 10 Hz

 

 

The top-view of the SF dataset sequences superimposed on the satellite image of the fort
The top-view of the SF dataset sequences superimposed on the satellite image of the fort
A UAV collectinf data at SF
A UAV collectinf data at SF
The DJI Mavic Pro
The DJI Mavic Pro

Leave a Reply

Your email address will not be published. Required fields are marked *