Profile cover photo
Profile photo
Niko Gamulin
162 followers -
Work for a cause, not for applause. Live life to express, not to impress.
Work for a cause, not for applause. Live life to express, not to impress.

162 followers
About
Communities and Collections
View all
Posts

0xB3C2E413F800D3030d089aE50819C2848eEbF0ed
Add a comment...

Post has attachment
Add a comment...

Post has attachment
Hello all!

As I am looking for ways to detect objects in real-time with mobile phone camera, I would be very thankful if anyone of you shared any experiences or comments about feasibility taking into account the computing power available on most commonly used mobile phones and the option of transmitting the stream into a cloud and returning results to mobile phone. I am a little skeptical about the latter option due to most likely latency. In case of processing stream on mobile device, unfortunately so far I have only managed to run OpenCV cascade classifier which performance is far worse than convolutional net. I have also tried to train convolutional net on server and deploy the model on mobile phone using caffe but in this case it was not possible to perform object detection in real-time.

Anyway, I would be glad to get some opinions about the way you tackled this challenge.

Hi all,

I've been thinking about using different corpus data for specific domain usages. 

If anyone is interested in cooperation, I've initialized a repository to generate corpus from subtitles files: https://github.com/nikogamulin/FoodCorpusComposer

Post has attachment

Is there any explanation why/when it is good to use convolutional neural networks with two or more consecutive convolution layers (like e.g. OxforfNet)? For some image recognition tasks I have used the architecture similar to some older models (convolution->nonlinear transformation->pooling) and it achieved better results than the network similar to OxfordNet. Anyway I am trying to find any resources which could help to improve reasoning about network architecture selection. I would be very thankful if anyone provided any hint or resource relevant to this question.

In general by testing a number of scenarios I have realized that the larger the number of layers (in cade of having a dataset large enough) the better the tesults are. Nevertheless there are a number of questions opened related to ar hitecture and sometimes I am wondering if the authors of the best models got to their results by reasoning or simply testing out a large number of different architectures.

Post has attachment
Add a comment...

Are there any general guidelines to design convolutional neural networks?

Specifically, it would be very helpful if anyone could answer any of the following questions:

Is there any rule to choose the right size and number of kernels given the properties of the input images (e.g. spectrograms, natural images, microscopic images...)?

How to select a good combination of kernel sizes over convolutional layers (is there any rule that suggest to increase/decrease or keep the same size of kernels on layers conv0, conv2, ..., convn)?

When is it reasonable to include two subsequent convolutional layers without pooling in between (like Alex Krizhsevsky did for ImageNet)?

Post has attachment
Add a comment...

Post has attachment
Add a comment...
Wait while more posts are being loaded