Feature extraction for THUMOS14 is strange

[makecent]

According to the "Implementation Details" part in Section 4 of the original paper, you use the model pre-trained on the training set of ActivityNet-1.3 as the features extractor. And I don't make anything wrong, your BMN uses the output of the last layer as the feature. That's why you get 400-dimensions length features (200 classes, 2 streams).

But the output of the last layer should represent the predicted class scores of the input frame of 200 ActivityNet actions. This kind of feature should be meaningless for THUMOS14 because half of the action classes in THUMOS14 don't belong to ActivityNet. How can you detect an unseen type of action using the classification scores of 200 irrelative actions?

[chenshen03]

the same problem

[xianguo-dev]

Maybe he means it is extracted after training with ActivityNet-1.3 pretrained model? It will make sense then.

[makecent]

Maybe he means it is extracted after training with ActivityNet-1.3 pretrained model? It will make sense then.

The feature extractor (pretrained model) is fixed without further training.