How do you test for action recognition?

[sukun1045]

What is the strategy during test step? Are you doing the forward propagation for all test sequences to extract the test features? Then how exactly to do the classification for action recognition problem? Thank you.

[xrenaa]

During the test step, in my understanding, we plot the training result and extract and plot the test feature and say the result. In the paper, it takes:

We follow the same evaluation protocol as defined in [36, 45]. All models are evaluated for the clustering quality and false positive rate (FPR) on the same test set which consists of unseen motion categories. We compute the FPR for 90%, 80% and 70% true positive rates. In addition, we also use the Normalized Mutual Information measure (NMI) and F1score to measure the cluster quality where the NMI is the ratio between mutual information and sum of class and cluster labels entropies while the F1score is the harmonic mean of precision and recall.

[36]. Song, H.O., Xiang, Y., Jegelka, S., Savarese, S.: Deep metric learning via lifted structured feature embedding. In: Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on. pp. 4004–4012. IEEE (2016)

[sukun1045]

What do you mean by plotting the training results? In Song's paper, they use affinity propagation clustering on test set features.

[xrenaa]

Just like the picture is shown in README.me. And actually we don't test this re-implement version in a numerical way for we re-implement this paper for another task.

[sukun1045]

Oh I see. So the xaxis is the classes, what is the y axis?

[xrenaa]

y is the score from the Attentive LSTM network, for the author uses the L2 norm at the end so the score is just a single value.