search

xlliu7 / MUSES

[CVPR 2021] Multi-shot Temporal Event Localization: a Benchmark
https://songbai.site/muses/
55 stars 4 forks source link

question for proposal generation #20

Open Go-MinSeong opened 1 year ago

Go-MinSeong commented 1 year ago

Hi, I'm trying to apply your Muses model to the Finegym dataset.

When I want to train your model using another dataset, I realized that I need a list of 100 candidate proposals.

I am referring to the supplementary of the paper for the details of proposal extraction. I am trying to perform proposal generation. However, I noticed that the multi-stage cnn used to generate proposals on the Muses dataset generates two final outputs for linear classification, whereas the final outputs for linear classification presented in the MUSES paper are three.

Therefore, if I want to train a binary classifier as suggested in the MUSES paper, how can I prepare the label values? Also, what do the three final outputs mean?

I would be grateful if you could answer my questions.

Go-MinSeong commented 1 year ago

I have additional questions. If I understand correctly, the supplementary of the paper generates a fixed length for the proposal. But the 100 candidate proposals from that repository all have different lengths, so how did you apply this?