feature generation and MC^2

[CHNxindong]

Great work.
I have several questions:

1. What is the difference between feature generation for MNIST and MC^2 transform for CIFAR10.
2. Why conduct these different process for these two datasets. How big results gap if no these transforms.
3. If I have face datast with 225*225 resolution, what's your suggestion about the transform(feature generation or MC^2)

非常好的工作！
我有几个问题想咨询：

1. 为什么MNIST和CIFAR10用不同transform？
2. 这两种transform(feature generation, MC^2) 有什么优势吗，不加的话差别有多大？
3. 如果我有人脸数据集，请问使用哪种transform比较好？

[zhangsz1998]

Thank you for you attention.

1. For relatively complex datasets such as CIFAR10, features produced by MCR^2 have good subspace structures, thus the performance of subspace clustering algorithms will be significantly improved. For simple datasets such as MNIST, using features extracted by scattering transform (the algorithm used in feature_generation.py) can already achieve high accuracy, so there is no need to use MCR^2.
2. As mentioned above, MCR^2 can produce features that have good subspace structures, and therefore improve the performance of subspace clustering algorithms. However, MCR^2 is a complex algorithm to run.
3. I haven’t tried to use subspace clustering for face clustering task, but I think simple scattering transform might not work and MCR^2 is more promising.

感谢您的关注：

1. 对于CIFAR10这类较为复杂的自然场景数据集，使用MCR^2能够抽取具有子空间结构的特征，因此后续子空间聚类算法的性能会有显著提升。而MNIST这些相对简单的数据集使用scattering transform就能取得不错的精度(也就是feature_generation.py中使用的算法)，因此没有必要使用复杂的MCR^2进行特征抽取。
2. 对于复杂的数据集，MCR^2算法能够抽取适用于子空间聚类算法的特征，因此提升较大，但其本身也更加复杂。
3. 我没有类似的经验，但我认为对于真实场景的数据集来说，由于简单的特征抽取无法得到符合子空间分布的特征，我猜测MCR^2特征抽取的效果会更好。