Get more training data

[freedomofkeima]

The result are quite bad when the number of classes is increased to 100. However, more images still make quite a difference in detection quality. With 30 images, the overall accuracy is 60.3% while with around 40 images, the overall accuracy is 64.0%. For details, see here.

One of the idea is to add more training data from animation video (screen capture).

[freedomofkeima]

Other ideas:

• Stability training: https://arxiv.org/pdf/1604.04326.pdf
• Fine-tune the last convolutional block: https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html
• Features selection improvement (not limited to face detection only)