Few inquiries regarding the content of your paper.

[gracikk-ds]

You've done an outstanding job!

I have a few questions about creating the ground truth for the dataset. You mentioned that the movies and trailers were preprocessed using shot segmentation and scene segmentation models. So, currently, we have scene timings for both the trailers and movies. Now, we need to find a way to align the moments in the trailers with the moments in the corresponding movies. You mentioned that you are using Faiss algorithms to measure visual similarity between frames in the trailers and frames in the movies. Does this mean that you are working at the frame level rather than the scene level? Are you extracting embeddings for each frame in both the movies and trailers and then matching them?

Could you please provide a detailed description of the creating the ground truth for the dataset process so that I can understand it better?