Proper Documentation about bhdd

[hectorhui]

Can you share you result for training? This look amazing, I am trying to do one Burmese digit recognition system (for research purpose only) with your data. I just want to compare and we can share some insight after I've done my model training.

[swanhtet1992]

@hectorhui
We have the plan to add benchmarks page in this repo. All of us are quite busy at the moment to get back to this. You are welcome to contribute the benchmark page.

There are a few people who built an app example and write an article with this data. Probably, you can reference their works.

We also have plan to collect additional raw data for the test set. If you are interested, let me know.

---

If you use BHDD or parts of BHDD for research, please cite this repo even though we haven't published our own paper for data processing.

[hectorhui]

@swanhtet1992
Thank you for your resources. I am building a CNN for BHDD dataset atm. I will share you my results and findings. May I ask how do you split the train set and test set?
From training result and testing result, i could get up to 99.8% accuracy, but anyway this sound too good to be true as well. I have not tested with any other hand written digits yet. I will try and let you know after my attempts.

[swanhtet1992]

May I ask how do you split the train set and test set?

We had 100+ contributors. We took test data from approximately 30 different writers.
Like I said in the above comment, we are planning to collect more data for test set (and probably for training set too).

The weakness in this dataset is that almost all contributors are at the same age. We noticed that hand writings can be pretty much the same for the age group. We are planning to add below 10 years old and above 40 years old age group. If you have sources where we can get contributors from such age groups, do let me know.

From training result and testing result, i could get up to 99.8% accuracy, but anyway this sound too good to be true as well.

That depend on your architecture and how deep you go. Since you are using CNN, it's kinda overkill for these kinds of datasets. Even for MNIST, LeNet-5 could achieve 0.8 test error rate. So, that's normal to have such high accuracy.

May be you could try Linear Classifier, KNN and SVM too.

[ThuraAung1601]

the pickel file is corrupted i think. i got this error while trying to is to continue my experiments.
UnpicklingError Traceback (most recent call last)
in ()
5
6 with open("/content/BHDD/data.pkl",'rb') as file:
----> 7 dataset = pickle.load(file)
8
9 trainDataset = dataset["trainDataset"]

UnpicklingError: invalid load key, 'v'.