search

keras-team / keras-cv

Industry-strength Computer Vision workflows with Keras
Other
1.01k stars 330 forks source link

Anchor free models #1

Closed bhack closed 4 years ago

bhack commented 4 years ago

Please evaluate to extract common components/utils that could support Anchor free models: https://github.com/tensorflow/hub/issues/424

See also: https://github.com/google/mediapipe/issues/495#issuecomment-616721458 https://github.com/google-coral/edgetpu/issues/51

MatthAmoros commented 4 years ago

Very naive questions here: 1) Why would you like to implement an anchor free model in an industrial context ? AFAIK anchor model are widely used for label inspection as they give a very quick and reliable result. 2) Have you considered adressing vision inspection problems (detect defects on products) or labeling only (read and check label information) ?

I would be very happy to discuss further on this subject, and maybe try to participate on this project, as I m currently trying to implement a (very humble) anchor based "old school" vision control library.

tanzhenyu commented 4 years ago

Anchor free models are probably not in our immediate roadmaps yet. @fchollet WDYT?

bhack commented 4 years ago

@tanzhenyu At least it would have a bit of a way to diversify keras-cv compared to the model garden since almost all official anchor free papers implementations are in pytorch.

EDIT: I meant that by an historical point of view model garden was more related to Google own research papers that on third_party research.

tanzhenyu commented 4 years ago

@tanzhenyu At least it would have a bit of a way to diversify keras-cv compared to the model garden since almost all official anchor free papers implementations are in pytorch.

EDIT: I meant that by an historical point of view model garden was more related to Google own research papers that on third_party research.

True -- but we'll focus on things not TF2/Keras implemented yet, such as semantic segmentation, DeepLab, etc. If this is in the long run outperforming anchor-based models, (as well as DETR I guess), then we'd need to consider having re-useable layers for them.

So "long run" is the vague word here, would it make more sense to put in addons?

bhack commented 4 years ago

but we'll focus on things not TF2/Keras implemented yet, such as semantic segmentation, DeepLab, etc

If you are talking about official models is true. But there are many third_party very good impl related to your list on Github.

tanzhenyu commented 4 years ago

but we'll focus on things not TF2/Keras implemented yet, such as semantic segmentation, DeepLab, etc

If you are talking about official models is true. But there are many third_party very good impl related to your list on Github.

Yes -- and I'd like to consolidate the efforts. If that's already very good impl we should ask the author to migrate directly to this repo (or Keras-applications if it's backbone)

bhack commented 4 years ago

@tanzhenyu nevermind anchor-free just appeared in the new model garden object detection API (Centernet) https://github.com/tensorflow/models/tree/master/research/object_detection#tensorflow-2-support

bhack commented 4 years ago

IMHO less reusable when we have utils in models repos https://github.com/tensorflow/models/blob/master/research/object_detection/utils/target_assigner_utils.py#L44