Ground Truth

I created some models for my testing set, rendered them in 3DS Max, and put them through the corner detector. Here are a few screenshots:

Rubik's


Earth


(The texture for the earth was taken from a satellite mercator projection I screencap-ed from Google Maps.)

Even among these simple test images, there is a large variation in the feature points that are selected. The lighting/shadows also seem to have a big impact. It looks like I still have a bit of tweaking to do with the parameters.

I thought it would be good to start with textured primitives like a cube and a sphere, but I will add more complicated objects too. The Multi-View Stereo Evaluation site also has some datasets that I can use.