Dynamic recalibration of stereoscopic camera system

One or more memories having contents configured to cause a computing system to perform a method, the method comprising:

• for each of a plurality of visual features depicted in both of two 2D images captured as a pair by a stereoscopic camera system, in each of the 2D images, determining a distance between an original position of the visual feature in the 2D image, and a second position of the visual feature in the 2D image obtained by (1) applying a computer vision model using a first set of calibration values to forward-project the original position of the visual feature in both of the 2D images to a projected position in 3D space, then (2) applying the computer vision model using the first set of calibration values to backward-project the projected position in 3D space back into the 2D images;
• determining a first error measure based upon the determined distances;
• applying an optimization mechanism to the first set of calibration values to obtain a second set of calibration values that seeks to reduce the determined first error measure;
• and
• storing the second set of calibration values for use with the stereoscopic camera system.

A method in a computing system, comprising:

• (a) accessing a first stereoscopic pair of 2D images of a first scene captured by a particular stereoscopic camera system, the first pair of 2D images including a left 2D image and a right 2D image;
• for each of the 2D images of the first pair:
• (b) automatically identifying a plurality of visual features occurring in the 2D image;
• (c) for each identified visual feature, determining a location of the visual feature in the 2D image;
• (d) matching visual features identified in the left 2D image with visual features identified in the right 2D image;
• for each visual feature identified in the left 2D image matched with a visual feature identified in the right 2D image:
• (e) applying a computer vision model using a first set of values for a plurality of calibration variables to forward-project locations of the matching visual features in the left and right 2D images of the first pair to obtain a 3D location for the matched features;
• (f) applying the computer vision model using the first set of values for the plurality of calibration variables to backward-project the obtained 3D location for the matched features to obtain round-trip-projected locations for the matched features in the left and right 2D images;
• (g) determining an error measure attributed to the first set of values for the plurality of calibration variables based upon, for each of at least a portion of the matched features, for both the left and right 2D images of the first pair, the distance between the matched feature's determined and round-trip-projected locations in the 2D image;
• and
• (h) applying an optimization mechanism to adjust the first set of values for the plurality of calibration variables to obtain a second set of values for the plurality of calibration variables, wherein the optimization mechanism seeks in making the adjustment to reduce the determined error measure.