Object detection: combined/seperate from the blur api?

I was wondering what the plan was for the object detection.
From my understanding YOLO which is currently used for the blurring does many detections in one pass.
So is the plan to have object detections also happen in the same time that the blurring detections happen? or will these be done separately in separate containers?

EDIT: is the plan to use the standard yoloV6 model for blurring but then custom models for other object detections? (and will these be done in a seperate container?)

There is no plan defined so far.

It is dependant a lot on the detection model. Currently the model detects 3 classes of object: faces, license plates and road signs.

We can train another model to include more classes, but :

  • it must not slow down the basic need which is blurring faces and license plates,
  • and must not lower the detection of these classes.

I think it will be a better approach to run additionnal object detection outside of the blurring pipeline, because we may do it also on limited areas or reduced picture resolution.

Hello Thibault,

We use Label Studio for detection via https://label-studio.cquest.org/
Face detection is already activated and blurring is automatic when you upload photos. If you want, you can have them blurred before you upload them. So a priori, there’s no need to do any more training. We tested blurring in Paris with different faces of tourists, but some may not come out as well (black people in particular).

During the 1st detection, we detected faces and licence plates for blurring as well as road signs for testing.

For the other objects, we’ll take a set of photos and detect a few classes of selected objects. Technically, there’s no limit to the number of object classes to be detected, but beyond 5, it seems complicated for the human mind, because you can’t forget any.
If we have more objects to detect, we’ll put the same photos back in for training, to make the job easier. Technically, you can detect a single class of object, but you have to go over each photo the same number of times.

The topic I opened on this subject: Nouvel entraînement de reconnaissance d’objets EP 2

Bonjour Thibault,

Nous utilisons Label Studio pour la détection via https://label-studio.cquest.org/
La détection des visages est déjà activée et le floutage est automatique quand on dépose les photos. Il y a possibilité de les faire flouter avant de les déposer si on veut. Donc a priori, pas besoin de refaire des entraînements. Nous avons testé le floutage à Paris avec différents visages avec les touristes mais il est possible que certains ressortent moins bien (personnes noires notamment).

Lors de la 1ère détection, nous avons détecter les visages et le plaques d’immatriculation pour le floutage ainsi que les panneaux de signalisation routière pour faire des tests.

Pour les autres objets, nous allons prendre un jeu de photos et détecter quelques classes d’objets sélectionnés. Techniquement, il n’y a pas de limite aux nombres de classes d’objets à détecter, mais au-delà de 5, cela semble compliqué pour l’esprit humain, car il ne faut pas en oublier.
Si nous avons plus d’objets à détecter, nous remettrons les même photos pour l’entraînement et ainsi se faciliter le travail. Techniquement, tu peux détecter une seule classe d’objets mais il faut repasser autant de fois sur chaque photo.

Le sujet que j’avais ouvert à ce sujet: Nouvel entraînement de reconnaissance d’objets EP 2

1 Like

That’s what I was hoping to hear. Also just because then you can have the gpu in the main server focus on blurring but maybe do object detection offsite (like a basement indeed xd)

1 Like

Panoramax goal in not to centralize things.

Think of it like OSM… the core of OSM is the database and main API, not much more.

Added value comes from the ecosystem around that core, not by waiting for the core to impliment one more nice feature.