Some attempts at OCR-ing the MNIST datasets.
The MNIST database site includes training & test data, plus a review of many techniques and their performance. LeCun 98 (pdf in papers/) also provides a broad overview of techniques.
To get the MNIST data files, run get-mnist-data.sh, or download the files
from the above site and use gunzip to decompress them.
- Python 3
- OpenCV (tested using 3.0.0)
- Numpy
- get memory usage info of classifiers
- parallelize knn classifiers
- predictall? should run faster
- renew benchmark figures of all classifiers
- test preprocessing stages separately, see what gets best results
- svm with pca. see https://github.com/grfiv/MNIST/blob/master/MNIST.pdf