EAST detect some texture as a text!

[hahmad2008]

I am running the demo using the pre-trained model

python eval.py --test_data_path=/tmp/images/ --gpu_list=0 --checkpoint_path=/tmp/east_icdar2015_resnet_v1_50_rbox/ \ --output_dir=/tmp/

However, EAST detect some features as text, is there any reason behind that? Is there any configuration to tune or change?

Samples:

[CarMiranda]

No, there is no "reason". EAST is just another deep learning model, so, as of right now, we cannot explain its outputs. And, again, as any other deep learning model, it can misbehave.

There are multiple things you can do:

• To improve accuracy in your particular data, you can tune the threshold parameters; maybe they are too low? This is not truly a recommendation since it does not improve your model...
• To improve overall robustness, train on more data, preferably adversarial to your case. Here, it predicts texture as text, so you can train on images with text and texture to force it to learn these kind of texture are not text. This is also known as hard example mining.
• To understand what is going on, try to interpret intermediate results (the score map f_score and geometry map f_geometry). Are intermediate results similar between these textures and text? If so, then the model has not learned to discriminate between them, and needs more training.

[francotheengineer]

@CarMiranda answer is correct.

You could also try using larger Backbones, more UNet layers, deeper box/score maps heads. Other mods could be changing activation functions, different Batch Norm layers. The list is endless.