Image retrieval web demo

[kloudkl]

To replace #243.

[kloudkl]

[1] Artem Babenko, Anton Slesarev, Alexandr Chigorin, Victor Lempitsky. Neural Codes for Image Retrieval. ECCV 2014.

[sunbaigui]

@kloudkl I've checked your commit code here, there's no sign you have implemented Neural Codes for Image Retrieval. If you have implemented this, could you kindly tell me where's it?

[kloudkl]

No, it has not been implemented.

That paper concludes that retraining the network on datasets closely related to the target one improves the image retrieval relevance and that the feature dimension can be greatly reduced with little performance loss.

Since this demo is based on the image classification demo of @sergeyk who has published a tutorial for fine-tuning the network on Flickr style dataset and a paper related to the above topics [2], I plan to re-use the image style classification model for searching. The sample images will come from the Flickr style dataset whose thumbnails are readily available online. The feature dimension will be reduced with PCA.

[2] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, Holger Winnemoeller. Recognizing Image Style. BMVC 2014.

[sergeyk]

Cool!

It so happens that I've recently implemented a demo myself:
http://similaritydemo.vislab.berkeleyvision.org
http://similaritydemo.vislab.berkeleyvision.org/similar_to/random/caffe%20fc6/euclidean

I will not be working on this more in the next few months. If you want to
look at the code, check out http://github.com/sergeyk/vislab in
vislab/UI/similarity.py and vislab/searchable_collection.py

http://similaritydemo.vislab.berkeleyvision.org/similar_to/random/caffe%20fc6/euclidean

On Tuesday, September 9, 2014, kloudkl notifications@github.com wrote:

No, it has not been implemented.

That paper concludes that retraining the network on datasets closely
related to the target one improves the image retrieval relevance and that
the feature dimension can be greatly reduced with little performance loss.

Since this demo is based on the image classification demo of @sergeyk
https://github.com/sergeyk who has published a tutorial for fine-tuning
the network on Flickr style dataset
https://github.com/BVLC/caffe/tree/dev/examples/finetune_flickr_style
and a paper related to the above topics [2], I plan to re-use the image
style classification model for searching. The sample images will come from
the Flickr style dataset whose thumbnails are readily available online
https://www.flickr.com/services/api/misc.urls.html. The feature
dimension will be reduced with PCA.

[2] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor
Darrell, Aaron Hertzmann, Holger Winnemoeller. Recognizing Image Style.
BMVC 2014.

—
Reply to this email directly or view it on GitHub
#1028 (comment).

[sergeyk]

Note that this dataset is only 80K images of the Flickr Style dataset. To
use larger dataset, you'll have to hash the feature values. @longjon has
done much work on this.

On Tue, Sep 9, 2014 at 7:56 AM, Sergey Karayev sergeykarayev@gmail.com
wrote:

Cool!

It so happens that I've recently implemented a demo myself:
http://similaritydemo.vislab.berkeleyvision.org
http://similaritydemo.vislab.berkeleyvision.org/similar_to/random/caffe%20fc6/euclidean

I will not be working on this more in the next few months. If you want to
look at the code, check out http://github.com/sergeyk/vislab in
vislab/UI/similarity.py and vislab/searchable_collection.py

http://similaritydemo.vislab.berkeleyvision.org/similar_to/random/caffe%20fc6/euclidean

On Tuesday, September 9, 2014, kloudkl notifications@github.com wrote:

No, it has not been implemented.

That paper concludes that retraining the network on datasets closely
related to the target one improves the image retrieval relevance and that
the feature dimension can be greatly reduced with little performance loss.

Since this demo is based on the image classification demo of @sergeyk
https://github.com/sergeyk who has published a tutorial for
fine-tuning the network on Flickr style dataset
https://github.com/BVLC/caffe/tree/dev/examples/finetune_flickr_style
and a paper related to the above topics [2], I plan to re-use the image
style classification model for searching. The sample images will come from
the Flickr style dataset whose thumbnails are readily available online
https://www.flickr.com/services/api/misc.urls.html. The feature
dimension will be reduced with PCA.

[2] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala,
Trevor Darrell, Aaron Hertzmann, Holger Winnemoeller. Recognizing Image
Style. BMVC 2014.

—
Reply to this email directly or view it on GitHub
#1028 (comment).

[kloudkl]

Recently a few new approximate nearest neighbor searching libraries further improved the state of art performance.

According to the benchmark of Wei Dong, KGraph was significantly faster than the widely used FLANN.

@maheshakya also conducted performance evaluation of several open source approximation nearest implementations Spotify ANNOY, FLANN, KGraph, nearpy and LSH Forest.

[maheshakya]

Hi @kloudkl , this shows the actual performances (precision against speed). But this has been done before I have implemented LSHForest in scikit-learn. I'm looking forward to do an overall evaluation after I complete LSHForest. You can get it from this branch. It has not been merged yet to scikit-learn master.

[shelhamer]

Closing since this preliminary PR was cancelled by the contributing fork and Sergey's image retrieval demo is online and the code is open source.