Dmytro Mishkin

This paper learns deep local patch descriptor (for replacing SIFT) by hard negative mining using current mini-batch. It outperforms SIFT and deep competitors on Oxford5K and Paris6K retrieval datasets.

Authors test different variant of CNN architectures, non-linearities, poolings, etc. on ImageNet.

Summary:
-  use ELU non-linearity without batchnorm or ReLU with it.
-  apply a learned colorspace transformation of RGB (2 layers of 1x1 convolution ).
-  use the linear learning rate decay policy.
-  use a sum of the average and max pooling layers.
-  use mini-batch size around 128 or 256. If this is too big for your GPU,
decrease the learning rate proportionally to the batch size.
- use fully-connected layers as convolutional and average the predictions for
the final decision.
- when investing in increasing training set size, check if a plateau has not
been reach.
- cleanliness of the data is more important then the size.
- if you cannot increase the input image size, reduce the stride in the consequent
layers, it has roughly the same effect.
- if your network has a complex and highly optimized architecture, like e.g.
GoogLeNet, be careful with modifications.

- SIFT family is still the best local descriptor, outperforms novel CNN [SiamNet2015] approaches.
- (adaptive) Hessian-Affine is the best detector with broad applicability (not beaten yet)
- Affine view synthesis greatly helps for non-geometrical problems.
- Datasets and WxBS-Matcher available http://cmp.felk.cvut.cz/wbs/
-  We need more diverse datasets for learning local descriptors than Yosemite and Libert 

For robust wide baseline matching:

1) Use combination of MSER and Hessian-Affine with RootSIFT as a descriptor

2) Do iteratively increasing affine view synthesis  - from sparse to dense

So you can match both fast for easy pairs and reliably for extreme (80 degrees of view point difference) pairs of same view of the object. Works for non-planar objects as well, much better than ASIFT.

Mean(input) = 0, var(input) =1 is good for learning. Independent input features are good for learning.
So:

1) Pre-Initialize network weights with (approximate) orthonormal matrices

2) Do forward pass with mini-batch

3) Divide layer weights by $\sqrt{var(Output)}$

4) PROFIT!