Computer vision library CCV 0.6 with new image classifier

For the free cross-platform computer vision library CCV, a new image classifier has been developed that has been trained in a convolutional neural network. For the first time a classifier of this level and models (face detector, car detector, pedestrian detector) are released under a free license.

One of the developers of CCV, a Chinese programmer, Liu Liu (Liu Liu), writes that over the past year and a half, there has been a real breakthrough in deep learning algorithms, and especially in the field of image classification based on convolutional neural networks. “For a library designed to work in the most modern implementations, it would be a shame not to implement a competent image classifier after more than a year of breakthrough research published on this topic,” he says. Moreover, among the free libraries there are quite a few implementations of such a classifier: Caffe , OverFeat and cuda-convnet , and besides, they are all under scientific licenses. Of course, the situation needed to be corrected.

As a result, Liu Liu spent five months - and nevertheless made a qualifier of decent quality ccv_convnet.c , which is included in the package of the new version of CCV 0.6.
')
The classifier was trained on the ILSVRC 2010 data set from 1000 classes, it demonstrates the top-1 error probability of 36.83% and the top-5 error probability of 16.25%, which is close to the highest class proprietary classifiers (the world's best Clarifai classifier showed on the ILSV set 2013 probability of top-5 error ( 11.19% ).

The class factor and related models are published under the international license Creative Commons Attribution 4.0.

You can test the classifier at the DoComputersDream.org website and through the RESTful interface .