Keystep Recognition

Ego4D Challenges
1

Hi everyone. I am working on the keystep recognition benchmark. The Ego-Exo4D paper says: “The keysteps in our dataset exhibit a very longtailed distribution. To address this challenge, we set a cutoff threshold at 20 samples per key step, limiting our analysis to 278 unique key steps”. It raises a couple of questions:

  1. The limit cutoff of 20 is from the training set or considering train+val+test?
  2. Will the testing set of the challenge contain only the 278 key steps or all of them?
2

I have a similar question, please provide the list of keysteps used in the paper as a baseline. Since, we don’t have access to the test set, it is hard to figure out which keysteps were used in the paper.

1 Like
3

Hi @dpatel, it is yet unknown to me the actual keysteps. However, i was able to replicate the results of the paper with a cutoff of 14. Meanin, i dicsard the keysteps that have less than 14 samples in train+val. The total number of keysteps after the cutoff is 284 so 5 keysteps more than what is reported in the paper. If you increase the cutoff to 15 you will be left with less than 272 classes. I belive the cutoff of 20 used in the paper involves sample from all the splits i.e. 20 samples in train+val+test.

4

Thank you for the information @dikoanxh, it is very helpful to know you were able to replicate the results.

5

@dikoanxh @dpatel
Thank you for the question. Please see the keystep info here for the benchmark.

Please feel free to let me know if there is any further question!