Skip to content
This repository was archived by the owner on Jan 7, 2025. It is now read-only.

how to use annotation on sense studio? #131

Closed
thunder95 opened this issue Mar 9, 2021 · 4 comments
Closed

how to use annotation on sense studio? #131

thunder95 opened this issue Mar 9, 2021 · 4 comments
Assignees

Comments

@thunder95
Copy link

There are splitted video frames on website of sense video.

Any guidelines or tutoral for that?

Background is supposed to be no action clip.
whats the difference about category_tag1, category_tag2. Both of them distinguish the start and end of each action?

@thunder95 thunder95 changed the title how to use annotaion on sense studio? how to use annotation on sense studio? Mar 9, 2021
@corneliusboehm
Copy link
Contributor

corneliusboehm commented Mar 9, 2021

Except for this Wiki Page we don't have much documentation on the annotation tool yet, because we're still changing a lot there, sorry.

The general idea is that you can tag up to two specific events in your videos, which would be your category_tag1 and category_tag2. For example when doing squats, you could annotate the low/sitting and the high/standing position. This would enable the model to detect those and even count how many squats you are doing by checking how often you alternate between those two positions.

In SenseStudio, go through each video and click the corresponding tag buttons on the frames that you want to annotate, then click Submit. Once all your videos are annotated, you can train your classifier again, but this time make sure to include the --temporal_training argument.

@thunder95
Copy link
Author

very nice idea, this annotation paradigm shows me a brand new way to handle with my problem (specifically, periodic motions). A more question w.r.t. this issue, how to deal with the middle position, for instance, the middle position between high and low, or left and right. Should I take it as background? @corneliusboehm

@corneliusboehm corneliusboehm self-assigned this Mar 10, 2021
@corneliusboehm
Copy link
Contributor

As long as you don't need the model to specifically react to the middle position, you can leave it as background. It's usually enough to just tag the peaks of the motion.
In the future we will also add support for adding more than two tags to a class so that even more complex sequences can be annotated.

@thunder95
Copy link
Author

ok, jetzt mir is alles klar. thank you!

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants