Do the Experiments on the Cohn-Kanade dataset. And I only use about 600 images (nearly 500 images for training, 100 images for testing, 12 AU, no alignment ).
Compare with and without Region Layer. In the situation of without Region Layer, I use one convolution layer to replace it.
Directly train without sample operation to deal with imbalance between positive and negative samples. So the dataset only contains label (1, -1)
Calculate loss according to the formula in Paper which considers the label {-1, 0, 1}. So If you want to do the paper's experiments (positive and negative samples for each AU), you can rewrite the lib/data_loader.
Only calculate the F1-score.
You can see the results in log files
Visualization The result with region layer is worse than without region layer. I think it maybe have something to do with
Compare to the results in paper (Some AU is different from the AU in my experiment)