action-recognition-0001-encoder embeddings

[m__akshaya]

Hi, 

Does action recognition encoder produce embedding for untrained actions? 

Will it work if we use these embedding for action matching similar to Facenet or Sphereface?

Thanks,

-a

[m__akshaya]

Hi,

Any updates?

Thanks,

Akshaya M

[Vadim_A_Intel]

Great question!

This model is trained with cross entropy loss (= softmax loss), as may you know from sphereface and similar papers this result in embeddings that are not discriminative enough (i.e. there are low intra-class distance and high inter-class distance between them) so theoretically this model is not suitable for this problem setting. But in practice this still might work with certain accuracy, although we have not done this kind of experiments. Anyways, the training code for this model is available at https://github.com/opencv/openvino_training_extensions/tree/develop/pytorch_toolkit/action_recognition so you can try to re-train it with appropriate loss (e.g. AM-SoftMax, Triplet, etc).

Tell us if you will find out something interesting!

--

Vadim

[m__akshaya]

Hi Vadim,

Thank you. I am interested in training with my own database. I hope your traning code is useable. 

Thanks,

Akshaya M

[Shubha_R_Intel]

Dear Vadim, 

Thank you for a thoughtful and succinct answer.

Shubha