I am reading a paper, In which they mentioned about
Each set of shared weights is called a kernel,
or a convolution kernel. Thus, a convolutional layer with n
kernels learns to detect n local features whose strength across
the input images is visible in the resulting n feature maps.
I don't understand the meaning of 'visible' in this context.
Generally speaking, each convolution kernel is responsible for a specific kind of pattern, e.g. edges of different direction, corners, blobs of different colors and etc. For example, if a convolution kernel is responsible for horizontal edges, then the resulting feature map will have evident activation in corresponding location. The strength of each kernel is visible in the resulting feature map because the patterns they are responsible for will have evident activations.
Here is a demo in a browser which you can play with -> https://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist.html
Related
Each layer in a CNN reduces the size of the input via convolution and max-pooling operations. Convolution is translation equivariant, but max-pooling is translation invariant. Correct me if this is wrong : each time max-pooling applied, the precise location of a feature is reduced. So the feature maps of the final conv layer in a very deep CNN will have a large receptive field (w.r.t the original image), but the location of this feature (in the original image) is not discernible from looking at this feature map alone.
If this is true, how can the accuracy of bounding boxes when we do localisation be so good with a deep CNN? I understand how classification works, but making accurate bounding box predictions is confusing me.
Perhaps a toy example will clarify my confusion;
Say we have a dataset of images with dimension 256x256x1, and we want to predict whether a cat is present, and if so, where it is, so our target is something like [sigmoid_cat_present, cat_location].
Our vanilla CNN (let's assume something like VGG) will take in the image and transform it to something like 16x16x256 in the last convolutional layer. Each pixel in this final 16x16 feature map can be influenced by a much larger region in the original image. So if we determine a cat is present, how can the [cat_location] be refined to value more granular than this effective receptive field?
To add to your question - how about pixel perfect accuracy of segmentation boundary !!
Your intuition regarding down-sampling via max-pooling is correct. Normal CNNs have that limit. However, there have been some improvements recently to overcome it.
The breakthrough to this problem came in 2015-6 in the form of U-net and atrous/dilated convolution introduced in DeepLab.
Dilated convolutions or atrous convolutions, previously described for wavelet analysis without signal decimation, expands window size without increasing the number of weights by inserting zero-values into convolution kernels. Dilated convolutions have been shown to decrease blurring in semantic segmentation maps, and are purported to work at least in part by extracting long range information without the need for pooling.
Using U-Net architectures is another method that seeks to retain high spatial frequency information by directly adding skip connections between early and late layers. In other words, up-sampling followed by down-sampling.
In TensorFlow, atrous convolutions are implemented with function:
tf.nn.atrous_conv2d
There are many more methods and this is an ongoing research area.
Assume we have a neural network that gets two inputs. the first input is location and size of a object and the second one is an image of the object. the location and size go through an MLP that map 4 dimensional input to 512 dimensional vector and the image go through ResNet34 which gives us a 512 dimensional vector that describes appearance of the object. After obtaining them position vector and appearance vector are summed to obtain a singular vector. Then the vector goes through the rest of the network.
After training the network, I gained a bad accuracy. I analyzed what happens in the network, and I realized that position vector is not treated similarly as appearance vector and appearance branch has more weight in calculations.
I want my appearance and position features have the same impact. How should I achieve this?
Instead of summing your image vector with the vector from the additional features, I would suggest concatenating them - so you end up with a 1024 dimensional vector. Then the layers that come after this concatenation can determine the relative impact of the features through your loss function.
This should allow the model to rely more heavily on whatever features result in the lowest loss.
I'm trying to understand fast(er) RCNN and following are the questions I'm searching for:
To train, a FastRcnn model do we have to give bounding box
information in training phase.
If you have to give bonding box information then what's the role of
ROI layer.
Can we use a pre-trained model, which is only trained for classification, not
object detection and use it for Fast(er) RCNN's
Your answers:
1.- Yes.
2.- The ROI layer is used to produce a fixed-size vector from variable-sized images. This is performed by using max-pooling, but instead of using the typical n by n cells, the image is divided into n by n non-overlapping regions (which vary in size) and the maximum value in each region is output. The ROI layer also does the job of proyecting the bounding box in input space to the feature space.
3.- Faster R-CNN MUST be used with a pretrained network (typically on ImageNet), it cannot be trained end-to-end. This might be a bit hidden in the paper but the authors do mention that they use features from a pretrained network (VGG, ResNet, Inception, etc).
So, I've seen that many of the first CNN examples in Machine Learning use the MNIST dataset. Each image there is 28x28, and so we know the shape of the input before hand. How would this be done for variable size input, let's say you have some images that are 56x56 and some 28x28.
I'm looking for a language and framework agnostic answer if possible or in tensorflow terms preferable
In some cases, resizing the images appropriately (for example to keep the aspectratio) will be sufficient. But, this can introduce distortion, and in case this is harmful, another solution is to use Spatial Pyramidal Pooling (SPP). The problem with different image sizes is that it produces layers of different sizes, for example, taking the features of the n-th layer of some network, you can end up with a featuremap of size 128*fw*fh where fw and fh vary depending on the size of the input example. What SPP does in order to alleviate this problem, is to turn this variable size feature map into a fix-length vector of features. It operates on different scales, by dividing the image into equal patches and performing maxpooling on them. I think this paper does a great job at explaining it. An example application can be seen here.
As a quick explanation, imagine you have a feature map of size k*fw*fh. You can consider it as k maps of the form
X Y
Z T
where each of the blocks are of size fw/2*fh/2. Now, performing maxpooling on each of those blocks separately gives you a vector of size 4, and therefore, you can grossly describe the k*fw*fh map as a k*4 fixed-size vector of features.
Now, call this fixed-size vector w and set it aside, and this time, consider the k*fw*fh featuremap as k featureplanes written as
A B C D
E F G H
I J K L
M N O P
and again, perform maxpooling separately on each block. So, using this, you obtain a more fine-grained representation, as a vector of length v=k*16.
Now, concatenating the two vectors u=[v;w] gives you a fixed-size representation. This is exaclty what a 2-scale SPP does (well, of course you can change the number/sizes of divisions).
Hope this helps.
When you use CNN for classification task, your network has two part:
Feature generator. Part generates feature map with size WF x HF and CF channels by image with size WI x HI and CI channels . The relation between image sizes and feature map size depends of structure your NN (for example, on amount of pooling layers and stride of them).
Classifier. Part solves the task of classification vectors with WF*HF*CF components into classes.
You can put image with different size into feature generator, and get feature map with different sizes. But classifier can only be training on some fixed lengths vectors. Therefore you obviously train your network for some fixed sizes of images. If you have images with different size you resize it to input size of network, or crop some part of image.
Another way described in the article
K. He, X. Zhang, S. Ren, J. Sun, "Spatial pyramid pooling in deep convolutional networks for visual recognition," arXiv:1406.4729 2014
Authors offered Spatial pyramid pooling, which solve the problem with different image on the input of CNN. But I don't sure is spatial pyramid pooling layer exists in tensorflow.
I recently found the "global_pooling" flag in the Pooling layer in caffe, however was unable to find sth about it in the documentation here (Layer Catalogue)
nor here (Pooling doxygen doc) .
Is there an easy forward examply explanation to this in comparison to the normal Pool-Layer behaviour?
With Global pooling reduces the dimensionality from 3D to 1D. Therefore Global pooling outputs 1 response for every feature map. This can be the maximum or the average or whatever other pooling operation you use.
It is often used at the end of the backend of a convolutional neural network to get a shape that works with dense layers. Therefore no flatten has to be applied.
Convolutions can work on any image input size (which is big enough). However, if you have a fully connected layer at the end, this layer needs a fixed input size. Hence the complete network needs a fixed image input size.
However, you can remove the fully connected layer and just work with convolutional layers. You can make a convolutional layer at the end which has the same number of filters as you have classes. But you want one value for each class which indicates the probability of that class. Hence you apply a pooling filter over the complete remaining feature map. This pooling is hence "global" as it always is as big as necessary. In contrast, usual pooling layers have a fixed size (e.g. of 2x2 or 3x3).
This is a general concept. You can also find global pooling in other libraries, e.g. Lasagne. If you want a good reference in literature, I recommend reading Network In Network.
We get only one value from entire feature map when we apply GP layer, in which kernel size is the h×w of the feature map. GP layers are used to reduce the spatial dimensions of a three-dimensional feature map. However, GP layers perform a more extreme type of dimensionality reduction, where a feature map with dimensions h×w×d is reduced in size to have dimensions 1×1×d. GP layers reduce each h×w feature map to a single number by simply taking the average of all hw values.
If you are looking for information regarding flags/parameters of caffe, it is best look them up in the comments of '$CAFFE_ROOT/src/caffe/proto/caffe.proto'.
For 'global_pooling' parameter the comment says:
// If global_pooling then it will pool over the size of the bottom by doing
// kernel_h = bottom->height and kernel_w = bottom->width
For more information about caffe layers, see this help pages.