Skip to content

Annotation database management and history after inference #1738

Description

@ltetrel

Is your feature request related to a problem? If so, Please describe.
Currently when a user loads an annotation file, there are two options:

  1. Either overwrite the existing annotations
  2. Or append new annotation to the current dataset (with an optional "prepend" type)

Then, if a user runs an inference pipeline, it currently force the overwrite of the existing groundtruth, which is not necessarily expected from a user. When sharing datasets, this can be confusing, were the annotations manually labeled, or were those detected ? What if I want to still keep the orginal groudtruth after loading another detection set ?

Another issue comes from the fact that the user does not have a choice when training a model: which annotations will be used?
If you run inference on an existing dataset, the newly created "annotations" (detections) will serve as groundtruth when we train a new model. This is not really expected, by default I want to train from a ground truth annotation, or maybe from generated annotations.

Describe the solution you'd like
Having a more elaborated versioning / history mechanism for annotations, and being able to select which version of annotation I want. Also decouple the notion of "ground-truth" annotations (for training and evaluation) from "detected" annotations (for evaluation, or production).

Describe alternatives you've considered
For decoupling, a simple solution would be to update the dataset with new annotation prepended by for e.g. "detected_".

Additional context
This ticket was opened so we can start discussing if that makes sense for dive

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions