samsunglabs / drop-dtw Goto Github PK

Dear, authors

I am very interested in your great Drop-DTW work.

May I ask when will you publish the code?

Many thanks!

Hi, great work,
But I run into a bit a problem while using drop dtw with manually set drop cost. Namely I have two sequences for different length, I found that no matter how I set the drop cost. nothing in the short sequence is dropped and each item in the shorter sequence is matched to exactly one item in the longer sequence.

Hey,

I was going over the Drop-DTW code and i found a small bug in the this line:

I believe this should be x_dropped = [N] if cur_state == 1 else []

Does the DTW algorithm work in blocks of 3 x K? where K are all the text embeddings, and we try and match it with 3 frame embeddings of the video in one go?

Thank you for sharing the code and appreciate the work.

I am writing to request your inputs on using Drop-DTW method for YouCook2 dataset towards step-localization task.
As YouCook2 Dataset comes with labeled captions for steps or video segments i.e. not discrete as in COIN/CrossTask. I am wondering how did you process them:

• Did you remove stop words and use S3D model to compute features?
• Did you use any other language model to compute the sentence features?
• Did you discretize the captions into labels, if yes how?
• Looking forward to your reply.

Thanks