Google Street View House Number(SVHN) Dataset, and classifying them through CNN
Google Street View House Number(SVHN) Dataset Link
Much similar to MNIST(images of cropped digits), but SVHN contains much more labeled data (over 600,000 images) with real world problems of recognizing digits and numbers in natural scene images.
Dataset is obtained from house numbers in Google Street View images.
Here we are classifying 32 x 32 cropped images given in format 2
Using CNN architecture.
data_preprocess.ipynb
: preprocess the data
svhn_model.ipynb
: run the model and report results
- Confusion metric
- Visualization of misclassified and classfied images
Findings:
From logs we can make out that dropout rate should be higher to learn
good features as images have lots of others digits image pixels also in it.
So it's get confused more often`
Note:
Above experiment is performed under 4GB RAM and 1GB GPU Memory
So it was difficult to train it for more steps, and adding more layers in architecture