About Me

I am a Research Associate in University of Cambridge, and a member of the Natural Language and Information Processing Group. My main focus is on the ALTA project, an industry collaboration with Cambridge English, with the goal of creating innovative new technologies for language learning and assessment. I’m also a visiting lecturer in Tartu University, teaching a course on machine learning and language modelling.

For more information, see my homepage: http://www.marekrei.com/

2 comments

  1. Saneesh

    Hello Marek,

    This email is to clarify some of my confusions and questions about deeplearning. So I decided to send to you.

    I came to DeepLearing to solve one of feature in my mobile app for fashion products. The idea is if you take a photo in your mobile the system should display what are the objects in that image, something like “round collor black tshirt with cartoon print” or “white and black check formal shirt full sleeve”. I saw some sites does this, but I’m not sure how much I can do to reach this level. At least I want to recognise the apparel, color and pattern. This will help the user to search without typing the details. At last the relevent images also need to list for the searched image.

    Current situation:
    I have successfully configured Ubuntu 14.04 with CUDA, caffe, NVIDIA DIGITS etc. Using DIGITS I created a database and trained the images for “ship/no ship” example using “jeans/tshirt”.

    Questions:
    1. For the above requirement the things I’m following is correct? or is there any better approch?
    2. If this is okay then how can I find the product with its color, patter etc.?
    3. What should be the folder structure to achieve this?
    4. Is there any existing database for fashion products? if so, how can I use it?

    I have PHP background, I didn’t understand completely about caffe, DIGITS etc. So if you can share some links which explains the things a little easy that would be great. (I’m sorry to say, http://caffe.berkeleyvision.org seems difficult to me!)

    I know, its big email with many questions, if possible please help!

    Regards,
    Saneesh.

    • Marek

      Hi Saneesh,

      You will need to train a classifier for each type of property you want to detect in the image, like colour, pattern, etc.
      And you will need manually annotated training data for these classes.

      I’m not sure there are many better resources available. I would suggest going through some Caffe tutorials on training new models first, like:
      http://caffe.berkeleyvision.org/gathered/examples/imagenet.html
      http://caffe.berkeleyvision.org/gathered/examples/mnist.html

      There is also some more recent work on image caption generation, which aim to generate a nice descriptive string for each image. Although I can’t predict how well that will work on your task. For example:
      Kiros, R., Salakhutdinov, R., & Zemel, R. S. (2014). Unifying Visual-Semantic Embeddings with Multimodal Neural Language Models.

Post a comment

You may use the following HTML:
<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>