Improve keypoints-with-depth transformation estimation algorithm

[integraledelebesgue]

Issues

• At the moment, implementation is very simple - we solve PnP problem between images with depths
• They should be properly transformed into pointclouds in 3D world coordinates
• Estimation algorithm is actually wrong - keypoints of actors are taken and processed naively - one of the side view perspectives is correct, but the opposite one gives wrong answer
• In fact, proper estimation without hardcoding actors order can be achieved only by matching faces (Issue #5)

Additional ideas

• Kabsch algorithm could be extended with colour matching... (not necessarily RGB colors, but maybe a hue or other representation)