Correspondence embeddings for stereoscopic reconstruction

Korespodencijska metrika slikovnih okana jedna je od metoda korištenih u stereoskopskoj rekonstrukciji. Stereoskopska rekonstrukcija je postupak analize slika iz dviju ili više konkurentnih kamera kojim se dobiva rekonstrukcija scene u 3D formatu iz nativnog 2D formata. Korespondencijska metrika slikovnih okana zasniva se na ugrađivanju okana u visokodimenzionalan prostor. Nakon ugrađivanja okana slijedi pronalaženje korespondentnih piksela te se kao izlaz dobiva mapa dispariteta. U okviru rada korišten je duboki model učen i testiran na javno dostupnom podatkovnom skupu. Rezultati eksperimenata vrednovani su izračunom rekonstrukcijske točnosti.The correspondence metric of image patches is one of the methods used in stereoscopic reconstruction. Stereoscopic reconstruction is the process of analyzing images from two or more concurrent cameras to obtain a reconstruction of a scene in 3D format from the native 2D format. The correspondence metric is based on the embedding image patch to a high-dimensional metric space. After such embedding, the next phase is finding correspondent pixels. Disparity map is obtained as an output of calculation the most similar pixels. In the paper there is a explanation of a deep model, which is learned and tested on a publicly available data set. The results of the experiments were evaluated by the calculation of the reconstructive accuracy