Code submitted in the iNaturalist Kaggle Competition

The code contains two different nets.

The first nets is the well-known inception-resnet model. The code allows to load pretrained weights from Imagenet, and perform transfer learning for use them in the new dataset.

The second model is a Harris corner detection model. It is a deep novel implementation of the famous Harris corner detections, where corner score maps have been blurred in order to obtain areas of maximum activation. Those maps are convoluted with normalized images with the objective of giving maximum values on those areas, it works as an attention map in the code.

Normalized images and Harris map images go through the inception-resnet model giving as a result the final features used for training.

A simple convolution layer is used for the final stage of the classification. From features map to a prediction vector.