Feat: Annotate Directory of Images

[JSabadin]

Purpose

This feature annotates a directory of images by performing predictions with a model and creating a dataset in LDF from those predictions.

A new command annotate is added to the CLI, and you can also use it via model.annotate(...).

The annotate functionality uses the config YAML file where the model is defined, and performs inference on the given image directory. For example, if a user defines a detection head like this:

nodes:
  ...
  - name: EfficientBBoxHead
    params:
      download_weights: False
      initialize_weights: True
      conf_thres: 0.03
      iou_thres: 0.65
    inputs:
      - RepPANNeck
  ...

The annotate process will use the conf_thres value of 0.03 and iou_thres of 0.65 when performing Non-Maximum Suppression (NMS) for that head.

By default, annotate supports all built-in luxonis-train tasks. If the head operates in a way that's aligned with these default tasks, it will work out of the box since the base head is designed for compatibility.

However, if custom post-processing of labels is needed, the user must implement their own annotate function for the head. In such a case, the user can apply any custom filtering logic to the model outputs.

Specification

None / not applicable

Dependencies & Potential Impact

None / not applicable

Deployment Plan

None / not applicable

Testing & Validation

None / not applicable

[tersekmatija]

Do we have some example usage?

Is it possible to configure also "post-processing" parameters like confidence, IoU threshold, ...? I assume that's defined through the architecture itself?

[JSabadin]

Do we have some example usage?

Is it possible to configure also "post-processing" parameters like confidence, IoU threshold, ...? I assume that's defined through the architecture itself?

@tersekmatija Model outputs can be grouped into common task types. That’s why there is a single annotate() function for the base head, which works with luxonis-train task types. For custom tasks, users can implement their own annotate() function for the head.

Regarding the parameters you mentioned: they are configured in the YAML config. However, for some custom parameters, users can modify their annotate() function to perform additional processing.