Different shape of inference features compared to soccernet features

[Daanfb]

Hi. I have a problem with the shape of features.

The Soccernet features I can download have a shape of (2700, 8576), but I do inference of a video I get a shape of (batch_size, 1, 768), being batch_size=448 for a video with duration 7:32 (25 fps).

How can I get the same shape?

[paulxinzhou]

@Daanfb The new features perform comparatively with the old features but are not a reproduction of the old features. You can run inference on the old videos and retrain the stage 2 network on the new features. Feel free to share your inference result with the new network and I'll try to find the features and the stage 2 network I ran. I didn't share them in the first place since the features were too big and I couldn't find a place to host them.

[weichen0407]

Sorry, what is the solution.

[paulxinzhou]

@weichen0407 See if this explains things. https://github.com/baidu-research/Soccernet-features/blob/develop/README.md#understanding-the-result-features