Model training#

You do not always need to train a model to use Annolid effectively. For many videos, you can label a single frame and track with a video-object-segmentation backend (e.g., Cutie / EfficientTAM-style trackers), then review and correct.

Train a model when you need:

  • higher speed for long recordings,

  • better generalisation to your camera/arena/species,

  • a domain-specific detector/segmenter (e.g., custom objects, special backgrounds),

  • a pose model with named keypoints.

Option B: Mask R-CNN via Detectron2 (optional)#

If you specifically need Detectron2-based Mask R-CNN training/inference:

  • Prefer the Colab notebook (File → Open in Colab) for a working GPU environment.

  • Or follow the Detectron2 installation guide linked from Install options.

COCO export#

If you need COCO format for interoperability with other toolchains:

  • GUI: File → COCO format

Note

Annolid can export multiple formats (COCO/YOLO/CSV). Pick the one that matches your training/inference stack.