Recalibrating the KITTI Dataset Camera Setup for Improved Odometry Accuracy
Abstract: Over the last decade, one of the most relevant public datasets for evaluating odometry accuracy is the KITTI dataset. Beside the quality and rich sensor setup, its success is also due to the online evaluation tool, which enables researchers to benchmark and compare algorithms. The results are evaluated on the test subset solely, without any knowledge about the ground truth, yielding unbiased, overfit free and therefore relevant validation for robot localization based on cameras, 3D laser or combination of both. However, as any sensor setup, it requires prior calibration and rectified stereo images are provided, introducing dependence on the default calibration parameters. Given that, a natural question arises if a better set of calibration parameters can be found that would yield higher odometry accuracy. In this paper, we propose a new approach for one shot calibration of the KITTI dataset multiple camera setup. The approach yields better calibration parameters, both in the sense of lower calibration reprojection errors and lower visual odometry error. We conducted experiments where we show for three different odometry algorithms, namely SOFT2, ORB-SLAM2 and VISO2, that odometry accuracy is significantly improved with the proposed calibration parameters. Moreover, our odometry, SOFT2, in conjunction with the proposed calibration method achieved the highest accuracy on the official KITTI scoreboard with 0.53% translational and 0.0009 deg/m rotational error, outperforming even 3D laser-based methods.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Explain it Like I'm 14
What is this paper about?
This paper looks at how the cameras in the popular KITTI dataset were calibrated and shows a new way to recalibrate them to get more accurate “odometry” — which means measuring how a car moves through the world using cameras. The authors design a one-shot (single photo) method to find better camera settings and prove that it makes several odometry algorithms more accurate, even beating some systems that use 3D laser scanners.
What questions did the researchers ask?
The team asked simple but important questions:
- Can we recalibrate the KITTI cameras using the provided calibration images to get better results?
- If we find better camera settings, will odometry algorithms (which estimate motion from images) become more accurate?
- Which camera settings matter most for improving accuracy, and how should we tune them?
How did they do the research?
The researchers focused on how to “calibrate” cameras — like making sure a camera has the right “glasses” so it sees the world correctly.
Key ideas explained simply
- Camera calibration: Setting the camera’s internal values (like focal length, the image center, lens distortion) and the positions of multiple cameras so the images match the real world. Think of it like adjusting a projector so the picture isn’t stretched or tilted.
- Stereo cameras: KITTI uses two cameras side-by-side, like your two eyes. The distance between them (the baseline) helps estimate depth.
- Odometry: Figuring out how the car moves (how far and how it turns) just from images.
- Reprojection error: Imagine marking the corners of checkerboard squares and predicting where they should appear in the image. The “reprojection error” is how far off your prediction is from the actual corner. Smaller error means better calibration.
What they changed compared to the original KITTI approach
- Original KITTI calibration used one photo with many checkerboard patterns. This is fast, but the boards in that single shot are far away, the squares look small, and lighting can be uneven. That makes corners harder to detect precisely.
- Instead of detecting corners directly, the authors first detect edges (the lines around squares) very accurately, then compute corner positions by intersecting those lines. Think of tracing the edges of tiles, and finding corners where two edges meet — this is often more precise than guessing the corner from pixels alone.
- After matching boards between the left and right camera, they optimize the camera parameters to minimize reprojection error (using a math method that adjusts settings until the predictions match the measurements well).
Tuning the most important settings
They noticed some camera settings are not tightly “locked down” by the single-shot boards — changing them slightly barely affects reprojection error, but can strongly affect odometry accuracy. So they did a careful “grid search,” which means:
- Try nearby values for key settings, check which ones make the car’s rotation and movement estimates most accurate, and pick those.
They focused on four settings:
- Left camera focal length in x (): like how “zoomed” the camera is horizontally.
- Left camera principal point (): the center of the image where the camera thinks “straight ahead” is.
- Stereo baseline: the distance between the two cameras (like the distance between your eyes).
They kept other parameters free to adjust during optimization.
What did they find and why does it matter?
- Their corner-by-line-intersection method gave smaller reprojection errors than the original KITTI method and OpenCV’s standard approach — meaning the calibration is more precise.
- When they tuned the four key settings with grid search, odometry accuracy improved a lot:
- About 30% better in translation (how far the car moves),
- About 50% better in rotation (how much it turns), on average across many sequences.
- They tested three different odometry algorithms (SOFT2, ORB‑SLAM2, and VISO2), and all got better with the new calibration.
- Their SOFT2 algorithm, combined with the new calibration, reached top performance on the official KITTI scoreboard:
- 0.53% translational error,
- 0.0009 degrees per meter rotational error,
- Even outperforming some methods that use 3D lasers.
This matters because KITTI is a widely used benchmark for self-driving research. Better camera calibration makes results more reliable and helps many algorithms perform better — not just one specific method.
What could this mean in the future?
- Better calibration can significantly improve odometry and SLAM (mapping while localizing), even if the dataset uses fixed camera images.
- The idea of tuning a small number of sensitive parameters (like focal length, principal point, and baseline) with a smart search could help other datasets and sensor setups too.
- Since KITTI is a long-lasting standard for evaluating algorithms, this work can lift the overall accuracy of camera-based systems and make comparisons fairer.
- In practice, teams working on self-driving and robotics can get more out of the same data by calibrating carefully — sometimes a small tweak in camera settings can give big gains in how well a vehicle estimates its movement.
Collections
Sign up for free to add this paper to one or more collections.