The purpose of this document is to help customers run Hammerhead on their cameras successfully.

Depth map is noisy

The depth map looks noisy, especially in the sky or uniform areas (e.g., white walls). Perhaps the depth maps look noisy compared to the HDK. The problems are ordered below from most to least common reasons.

1. Intrinsic camera calibration.
   1. Redo the intrinsic camera calibration. Re-projection error should be less than around 0.4 pixels. Do you get the same intrinsic parameters?
   2. The supported camera models are OpenCV pinhole and fisheye models. Check that you are using the right model.
   3. Still need help —> Ask for a premium support. Camera tuning service. You can send us the cameras to calibrate using our Image Engineering Geocal test equipment.
2. Extrinsic camera calibration.
   1. Is the initial guess of the extrinsics within 3 deg of the actual mount? Guess a different extrinsic.
   2. Still need help —> Ask for support. We can manually run extrinsic search over larger search range.
3. Camera image quality
   1. Highly compressed images with compression artifacts? For example, jpeg compression has periodicity.
   2. Enough light? Grainy images? Do higher SNR images improve results?
   3. Blurry due to motion artifacts? —> try static scene to see if results improve.
   4. Low contrast from dust on lens, rain, fog? —> Clean lenses.
   5. Still need help —> Ask for support. We can analyze your images for anomalies.
4. Calibration scene
   1. disparity search range. If scene is out of the search range, it will not converge. Most of the scene should be between 0-255 pixels of disparity. If something is too close, then it will not match.
   2. Too much sky. Mostly a white wall. Not much variation in the image. Not enough texture in the scene for calibration.
   3. Autocalibration can diverge if it is. It throwing out bad scenes. “Bad” was defined from automotive scenes and this may need tweaking for other situations (sea).
5. Camera issues
   1. Simultaneous capture. If the scene is moving, for example on a rocking boat, the left and right cameras will take images at different orientations if there is a delay in capture. Check that cameras are synced. One way, is to point both cameras at a camera monitor with a stop watch displayed on the screen (good to a few milliseconds). To get better measurements, one could use Image Engineering’s LED array device.
   2. Point with large overlap for default factory calibration algorithm. If you need something else, that would be custom work.