High precision low baseline stereo vision algorithms

The first part was carried out within the framework of the project DGA-ANR-ASTRID "STEREO". The aim is to define the effective limits of stereo reconstruction when controlling the entire acquisition chain at the maximum precision, that one acquires stereo pairs in very low baseline and noise-free.

To validate this concept, we create very precise ground truths using a renderer. By keeping the rays computed during rendering, we have very dense information on the 3D scene. Thus we create occultation and disparity maps whose precision error is less than $ 10 ^ {- 6} $. We have made synthetic images available to the research community with an SNR greater than 500: a set of 66 stereo pairs whose $ B / H $ varies from $ 1 / 2500$ to $ 1 / 50$.

To evaluate stereo methods on this new type of data, we propose metrics computing the quality of the estimated disparity maps, combining the precision and the density of the points whose relative error is less than a certain threshold.

We evaluate several algorithms representative of the state of the art, on the pairs thus created and on the Middlebury pairs, up to their operating limits. We confirm by these analyzes that the theoretical assumptions about the merit of the low $ B / H $ in high SNR are valid, up to a certain limit that we characterize.

We thus discover that simple optical flow methods for stereo matching become more efficient than more sophisticated discrete variational methods. This conclusion, however, is only valid for high signal-to-noise ratios. The use of the dense data allows us to complete the ground truths a subpixel detection of the occlusion edges. We propose a method to compute subpixel vector contours from a very dense cloud of points, based on pixel classification extit {a contrario} methods.

The second part of the thesis is devoted to an application of the subpixelian optical flow and extit {a contrario} methods to detect clouds in satellite  imagery. We propose a method that exploits only visible optical information. It is based on the temporal redundancy obtained by the repeated passages of the satellites over the same geographical zones.

We define four clues to separate the clouds from the landscape: the apparent inter-channel movement, Local texture, temporal emergence and luminance. These indices are modeled in the statistical framework of extit {a contrario} methods which produce an NFA (number of false alarms for each). We propose a method for combining these indices and computing a much more discriminating NFA. We compare the estimated cloud maps to annotated ground truths and the cloud maps produced by the algorithms related to the Landsat-8 and Sentinel-2 satellites. We show that the detection and false alarms scores are higher than those obtained with these algorithms, which however use a dozen multi-spectral bands.