The Street View House Numbers Dataset from Stanford has been used for this project. [http://ufldl.stanford.edu/housenumbers/]
The previous project(mnist dataset) had black and white images. Here 3 different colour channels are introduced in the data set, giving each image the size of 3x32x32 pixels. Again, I've written the code in both theano and tensorflow for learning purpose.
I created a model with 2 convolution-Pooling Layers plus a fully connected ANN. The lowest test error rate observed is 0.145.