Mar 2019
tl;dr: KITTI vision benchmark suite, and data preprocessing.
Overall impression
The KITTI dataset underwent quite a lot of preprocessing, including rectification (for stereo vision tasks), calibration, synchronization.
Key ideas
- Equipment
- 2 pairs of color/grayscale camera, 54 cm apart. It operates at 10 Hz, triggered when lidar heads forward.
- Velodyne HDL-64E (10 Hz), 2 cm distance accuracy, 0.1 degree angular resolution.
- GPS/IMU localization unit with RTK correction
- Tasks
- Stereo matching and optical flow: images are rectified from 1392x512 to 1240x376.
- 3D visual odometry/SLAM: evaluating 6D poses. For a given set of image pairs, compute the relative difference of the GT difference and prediction difference of 6D poses.
- 3D vision: annotate tracklets in 3D space and project them back to image. Displays lidar point cloud as well as the camera images to increase the quality of annotation.
- A subset of data is selected to evaluate each task.
- For 3D object detection, images are iteratively added to the subset according to non-occluded objects as well as entropy of the object orientation distribution.
- Image form one sequence does not appear in both training and test set.
Technical details
- GNSS (GPS, GLONASS, Beidou, Galileo), IMU and RTK correction signal. RTK (real time kinematics) is using the phase of the carrier wave to correct location signal and has a localization error of 5cm. The GNSS signal itself without RTK correction can only localize to a few meters.
- Camera images are calibrated with intrinsic and extrinsic parameters, with checkerboard image mounted on walls.
- Velodyne to camera calibration is done semi-manually.
Notes
- Uber has visualization tool AVS, Open Standard for Autonomous Vehicle Visualization.
Camera calibration
- Pinhole camera model is described in openCV documentation and P371 of Learning OpenCV by Gary Bradski.
- Note how the image plane (projective plane) is mathematically (though not physically manufacturable) reformulated to be in front of the camera. The pinhole (usually the center of the lens) is reinterpreted as the center of projection.
- Principal point in a pinhole camera model: intersection of optical axis on the image plane. Lens distortion calibration: may be separated from remaining calibration processes.
- Homogeneous coordinates (齐次坐标, or projective coordinates 投影坐标) is an extension of Euclidean coordinates and used in projective geometry. Any points whose coordinates are proportional are equivalent.
- Lens Distortions corrects for the deviation caused by the introduction of lens to the pinhole model. It consist of mostly radial distortion (from the shape of the lens) and slight tangential distortion (form assembly of the lens). It is usually determined by a distortion coefficient matrix of shape (1x5), k1, k2, p1, p2, k3. ki characterizes radial distortion (typically k1 and k2 are enough for non-fisheye lens) and pi characterizes tangential distortion. This is the simplified rule where Pu = f(Pd, rd), and rd is the distance of Pd to the principal point. Radial Distortion can take on different forms: barrel (for wide angle lens), pincushion (for telephoto lens), or complex mixture of the two.
- Intrinsic parameters have 4 DoF, principal points (cx, cy) and focal length (fx, fy). We need to know the location of principal points as it is usually not in the center of the image. We also need fx and fy as the pixel size in a cheap sensor is not necessarily the same. The matrix of intrinsic parameters does not depend on the scene viewed. So, once estimated, it can be re-used. Usually it is written in a 3x3 camera matrix. In some context, intrinsic parameters also include distortion.
-
| Extrinsic parameters have only 6 DoF, but usually uses 12 parameters in a 3x4 matrix [R |
t]. |
- OpenCV calibration works with a predefined pattern on a plane (a chessborad pattern). Without lack of generality, we set Z = 0 for the object plane, then [x, y, 1]^T = M [r1, r2, t] [X, Y, 1]^T = H [X, Y, 1]^T. The H is a 3x3 matrix that maps coordinates in the object plane to image plane. The homography matrix describes the transformation from a coordinate in object plane to a coordinate in image plane.
- Generally we use more points than equations for calibration, thus overdetermined.
Projection and 3D vision
- Affine and perspective transformation transforms a list of points (or a whole image) from 2D to another 2D. (This is to be differentiated from perspective projection which projects a 3D point into 2D, along a set of projection lines which all meet at a point named the center of projection). The perspective transformation, which is a specific kind of homography, relates two different images that are alternative projections of the same three-dimensional object onto two different projective planes.
- A 3D rigid body transformation has 6DoF and is called a pose. A pose belongs to special euclidean group SE(3). For common representations of a pose, see review paper here.
- Image rectification: image processing so that the epipolar lines are horizontal (i.e., the corresponding points can only be in the same row in the other image). After rectification, the images are as if taken from two co-planar cameras. See example image.
- Epipolar line: in stereo vision, the corresponding point in the other image can only lie in a line, this line is named epipolar line. If two images are coplanar, i.e. they were taken such that the right camera is only offset horizontally compared to the left camera (not being moved towards the object or rotated), then each pixel’s epipolar line is horizontal and at the same vertical position as that pixel. However, in general settings (the camera did move towards the object or rotate) the epipolar lines are slanted.
- 3D stereo
- Undistortion
- Rectification
- Correspondence and generation of disparity map.
- Depth map
