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.
- Total 10 Classes, 1 for each digits i.e Label '9' for digit 9 and '10' for digit 0.
- 73,257 digits for training, 26,032 digits for testing
- Available in two differnet formats
- Original images with bounding box available for each character (may contain multiple characters in same images).
- MNIST like 32x32 cropped images having single character in each 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