Learning a correspondence embedding for stereoscopic reconstruction

Stereoskopska rekonstrukcija postupak je kojim pomoću dviju kamera želimo rekonstruirati geometriju scene. Važan korak prije tog postupka jest odrediti korespondentne piksele. Korespondencija se može ostvariti pomoću ručno napravljenih metrika ili metrika naučenih dubokim neuronskim modelima. U radu je opisan postupak korespondencije koji ugrađuje slike u visokodimenzionalan metrički prostor, a zatim takva ugrađivanja koristi za pronalazak korespondentnih piksela tj. mape dispariteta. Preuzet je javno dostupan skup slika "KITTI 2015 stereo" te su iz tog skupa izvedeni podaci za učenje mreže. Provedeni su eksperimenti za vrednovanje točnosti rekonstrukcije na ispitnom skupu te opisani su koraci za daljni razvoj.Stereoscopic reconstruction is a procedure by which we want to reconstruct the geometry of a scene with the help of two cameras. An important step before this procedure is to determine the corresponding pixels. Correspondence can be achieved by using hand-crafted metrics or metrics learned by deep neural models. The paper describes the procedure of correspondence by which images are embedded in a high-dimensional metric space, and then such embedding are used to find the correponding pixels ie. disparity maps. A publicly available set of images "KITTI 2015 stereo" was downloaded. Dataset for learning neural network was derived. Experiments were conducted to evaluate the accuracy of the reconstruction at the test set. Steps for further development are described