We can see in this set of rescontruction images that the foot has some bad artifacts in it. The normals / depths were miscalculated.
It seems as though the reconstructions usually do not have smooth edges as the calculations tend to smear the geometry as the normals get closer to the edges of the mask.
This first set of images shows the trouble with hidden geometry very well. In the source images we cannot see behind the beak. When we rotate the reconstruction, it is evident that the algorithm made its best guess as to the geometry behind the beak. It also seems that the geometry was generated to find the shortest distance from edges of the whole created by ocluded geometry. We can see this without the albedo.
Not everything was bad though. The program seemed to do a good job of recovering and remapping color to the underside of the neck. This is a spot where the perspective from the source images caused the normals more perpindicular to the viewing angle, so I am surprised it is not very smeared.
The detail of this normal map is interesting.
With these two reconstructions we can see how the geometry and albedo seems to get smeared along the face where normals were more perpindicular to the view in the source images.
By moving the light around you can really start to see the shape accuracy of the
reconstructions.
Each part of the process worked well as compared to the sample solution, but if the measure of wellness is of the quality of reconstruction, then I would say that the program did not work so well with faces perpincular to the viewer and with edges along the border of the mask.
To improve on the reconstructions it might be handy to clamp values down. For example it looks as though depths near the edges of the mask are much deeper compared to neighbors, so it stretches the edge. If you could calculate the difference in depth between neighbors, it might be easy to control that. Or even just shrink the edge in and throw out data by the edge of the mask that you think is bad.
To fix the problems with hidden geometry and smearing of albedos it would be useful to increase your test data by creating different viewing angles and then combing those in the end.