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

wrong blurs #2652

Closed
jotpe opened this Issue Aug 7, 2017 · 7 comments

Comments

Projects
None yet
5 participants
@jotpe

jotpe commented Aug 7, 2017

I always correct the blurs of my sequences. The blur matches were improved some time ago, but after that step there was no noticeable improvement anymore.
I see a lot of false positives... Faces which are not blurred or license plate which are not or one letter is uncovered. Half faces or half license plates are never blurred.

Are corrections part of learning material? It is frustration to correct thousend of photos and to see every time the same time of errors.

I guess the blurring algorithm is independent from the new AI detection. From a end-users point of view I do not understand the bad quality of blurring. Because the new AI detection can recognize people and vehicles very well, this could be a good filter.

@jotpe

This comment has been minimized.

Show comment
Hide comment
@jotpe

jotpe Aug 7, 2017

This one f.e. has no blur, but I don't know why.
bildschirmfoto von 2017-08-07 18-37-47

jotpe commented Aug 7, 2017

This one f.e. has no blur, but I don't know why.
bildschirmfoto von 2017-08-07 18-37-47

@voschix

This comment has been minimized.

Show comment
Hide comment
@voschix

voschix Aug 7, 2017

voschix commented Aug 7, 2017

@ybkuang

This comment has been minimized.

Show comment
Hide comment
@ybkuang

ybkuang Aug 8, 2017

@jotpe @voschix We are currently working on improving the face/license plate detection. The corrections you've have made are essential to improve our algorithm.

In this new version of the detection algorithm, we're taking into account of context information e.g. locations of the pedestrians/vehicles (same as @jotpe suggested). In this way, we will ensure low false positives and low false negatives. With more training data, we will be able to also localize the face/license plate better, so we make sure that the full face/license plate will be blurred. In the meantime, we are finding a good balance between localizing all the small faces/license plates against the speed (cost) of the processing. To achieve that, we're making more engineering efforts.

Again, thanks for the valuable feedback and sorry for the delay on improvement.

ybkuang commented Aug 8, 2017

@jotpe @voschix We are currently working on improving the face/license plate detection. The corrections you've have made are essential to improve our algorithm.

In this new version of the detection algorithm, we're taking into account of context information e.g. locations of the pedestrians/vehicles (same as @jotpe suggested). In this way, we will ensure low false positives and low false negatives. With more training data, we will be able to also localize the face/license plate better, so we make sure that the full face/license plate will be blurred. In the meantime, we are finding a good balance between localizing all the small faces/license plates against the speed (cost) of the processing. To achieve that, we're making more engineering efforts.

Again, thanks for the valuable feedback and sorry for the delay on improvement.

@jotpe

This comment has been minimized.

Show comment
Hide comment
@jotpe

jotpe Aug 8, 2017

I think it can be a good idea for technical freaks, to see blurring statics in the weekly digest, too.

And I believe it would help the motivation of your blur correctors, to tell regularly about your work and status about the blurring algorithm. This allows the user to derive a relationship between his work and the algorithm quality.

jotpe commented Aug 8, 2017

I think it can be a good idea for technical freaks, to see blurring statics in the weekly digest, too.

And I believe it would help the motivation of your blur correctors, to tell regularly about your work and status about the blurring algorithm. This allows the user to derive a relationship between his work and the algorithm quality.

@andrewharvey

This comment has been minimized.

Show comment
Hide comment
@andrewharvey

andrewharvey Aug 28, 2017

And I believe it would help the motivation of your blur correctors, to tell regularly about your work and status about the blurring algorithm. This allows the user to derive a relationship between his work and the algorithm quality.

I agree, before reading this thread it's wasn't clear if manual blurs (both deleting false positives and adding false negatives) were actually fed back as training data.

andrewharvey commented Aug 28, 2017

And I believe it would help the motivation of your blur correctors, to tell regularly about your work and status about the blurring algorithm. This allows the user to derive a relationship between his work and the algorithm quality.

I agree, before reading this thread it's wasn't clear if manual blurs (both deleting false positives and adding false negatives) were actually fed back as training data.

@katrinhumal

This comment has been minimized.

Show comment
Hide comment
@katrinhumal

katrinhumal Sep 5, 2017

Contributor

@jotpe and @andrewharvey good point - we mention that manual blur fixes create data for improving the algorithm on our help pages: https://help.mapillary.com/hc/en-us/articles/115001663705-Blurring-images but will keep it in mind that we could do more explaining and updates around this.

Also, the "edits" that are counted in our weekly digest comprise both change edits (i.e. location, compass angle etc. map-related) and blur edits. But they are covered with one figure.

Contributor

katrinhumal commented Sep 5, 2017

@jotpe and @andrewharvey good point - we mention that manual blur fixes create data for improving the algorithm on our help pages: https://help.mapillary.com/hc/en-us/articles/115001663705-Blurring-images but will keep it in mind that we could do more explaining and updates around this.

Also, the "edits" that are counted in our weekly digest comprise both change edits (i.e. location, compass angle etc. map-related) and blur edits. But they are covered with one figure.

@ybkuang

This comment has been minimized.

Show comment
Hide comment
@ybkuang

ybkuang Jul 20, 2018

This should be improved by the recent release for privacy blurring. Please let us know if the issues persists.

You can see that the manual blurs help with the improvement of our algorithm.

ybkuang commented Jul 20, 2018

This should be improved by the recent release for privacy blurring. Please let us know if the issues persists.

You can see that the manual blurs help with the improvement of our algorithm.

@ybkuang ybkuang closed this Jul 20, 2018

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment