albertotono
commented
3 years ago In the paper they also report that it doesn't
"We note that using different shading coefficients or illumination models did not affect our output descriptors due to the invariance of the learned filters to illumination changes, as also observed in image-based CNNs [20, 10]."
I would also add that intra-class and inter class invariance in the dataset can influence the performance.
hello there, I generated my datasets based on your BlendPhong and modelnet40, finally got images like this.
white models in black background. But when I tested the cnn with my datasets, the accuracy is only 50%~60%.when I get rid of some classes familliar such as table&night_stand, the accuracy rose to 90%. I was confused. the difference between your datasets and my datasets is lighting and camera type(ORTHO&Perspective).Does these matter?