Code release for 360monodepth. With our framework we achieve monocular depth estimation for high resolution 360° images based on aligning and blending perspective depth maps.
This commit avoid nan values in the scale term 1/scale while also fixes some coordinates artifacts when projecting tangent images to ERP.
Focusing in the 1K input images (2048x1024) the north and south pole faces had vertical coordinate ranges [-1,x] and [x,1024] respectively. In the former that means the north pole faces have values at the south pole and vice versa, resulting in wrong blending weights. Originally this should be taken care of later in the code at gp.inside_polygon_2d but it is not working for some reason.
Also I add some small changes to speed up the computation of frustum blending weights. With more clean code.
This commit avoid nan values in the scale term
1/scalewhile also fixes some coordinates artifacts when projecting tangent images to ERP.Focusing in the 1K input images (2048x1024) the north and south pole faces had vertical coordinate ranges [-1,x] and [x,1024] respectively. In the former that means the north pole faces have values at the south pole and vice versa, resulting in wrong blending weights. Originally this should be taken care of later in the code at
gp.inside_polygon_2dbut it is not working for some reason.Also I add some small changes to speed up the computation of frustum blending weights. With more clean code.