This is a prototype.
This flask web app uses computer vision to automatically count wheat heads on
wheat photos. Manual wheat counting is tedious. This is an attempt
to solve that problem.
The two models that power this app were trained using data from the Kaggle
Global Wheat Detection competition.
The data is licenced under an open source MIT licence.
1-- This demo will be live until 30 June 2020.
2-- The models were trained using image augmentation to help them generalize i.e. produce accurate results irrespective of where in the world the photos were taken. However, this capability needs to be tested further.
3-- The models were trained on wheat images from the following regions:
usask_1, arvalis_1, arvalis_3, inrae_1, ethz_1, rres_1
If you submit competition dataset images that are from these
regions then the predicted wheat head count will be very accurate. This is
because the models memorized those images during training. A better way to test this app would be to submit images that are from the validation set, from the holdout set or not from any of the regions in the competition public dataset.
4-- Model performance was validated on photos from these regions:
The design code and the step-by-step process used to train and test the models has been published on Kaggle. You can find the open source notebook here. The test results are also available in the notebook.