Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

CVAT AI Tools: dextr usage #2515

Closed
2 tasks done
aschernov opened this issue Dec 1, 2020 · 0 comments · Fixed by #3473
Closed
2 tasks done

CVAT AI Tools: dextr usage #2515

aschernov opened this issue Dec 1, 2020 · 0 comments · Fixed by #3473
Assignees

Comments

@aschernov
Copy link
Contributor

My actions before raising this issue

Expected Behaviour

Current Behaviour

1. Annotation speed (manually annotation VS semi-automatic annotation using AI tools: dextr interactor
We made an experiment to compare manually polygon annotation speed and semi-automatic using the dextr interactor and got the following results:

Annotation type Total polygons Time Speed
manually 1335 11,2 119,2
dextr 1450 19,83 73,12

*dextr speed retard is 38,65%

2. DEXTR advantages and disadvantages
Advantages:

  • Accuracy regardless of zoom. It isn't necessary to use zoom to get a good annotation result:
    12(zoom)
    11(without zoom)

  • Works well on round-shaped objects located on a contrast background like these ones:

2

3

4

5

it took 5-10 seconds (put 4 points and wait a request to server) to annotate such objects using DEXTR
  • Works well enough on irregularly-shaped objects like trees, vegetation, bushes:
    6(7 sec, some corrections were made on the left side)
    7(9 sec, but 5 points)
    8(This one took 20 sec, but the result is impressive)
    9(10 sec, but 5 points)
    10(10 sec, but 5 points)

Disadvantages:

  • The user interface blocks every time during a server request. A user has to wait for 3-4 seconds with no opportunity to make something useful for annotation

  • Too many polygons points, especially for big objects. It affects CVAT performance and complicates corrections:

13

  • Often a completed polygon doesn't include a covered by initial points area:

14

  • Low annotation quality for long, thin, tall and strict-shaped objects:

16

17

To get this result with dextr we spent ~40 sec:

18

And ~ 12 sec to get the same result with manually annotation:

19

Possible Solution

3. Suggestions how to improve the tool

  • Add an opportunity to set a number of initial points. At the moment it's 4 by default but according to a user's experience he can make a conclusion that some objects annotate well enough if he will put, for example, 6 initial points and start the dextr extractor
  • Add a visual notification for users that they can press "ctrl" to enlarge number of initial points before and after creating a polygon by dextr
  • Reduce a number of points for polygons created by dextr
  • Make a request to server invisible for a user and avoid blocking user interface
  • Develop a multi-annotation mode for similar objects. For example, we see identical objects on a frame but they have different size. The idea is we could set initial points for these objects and start the dextr after that for all these objects at once.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging a pull request may close this issue.

2 participants