Ctrl+K

Technical overview

Technical overview#

At a high level Annolid’s workflow consists of the following steps:

  1. Labeling of frames (annotation) & COCO formatting

  2. Training (fine-tuning) and inference (local worstation or Google Colab)

  3. Post-processing and analysis

Overview of Annolid workflow

Accessibility and efficiency#

We work hard toward making Annolid as accessible as possible to anyone trying to use it and are striving to make all the code be also runnable on Google Colab.

  • Options for training on Google Colab as well as on a local workstation

  • Fast training with quality- and speed-optimized options

    • Model training :

      • 200 labeled images

      • < 2 hours for 3000 iterations on Google Colab

      • 30 min on NVidia 1080Ti

    • Inference (applying trained model to behavior videos)

      • Mask R-CNN: ~7 FPS

      • YOLACT: ~30 FPS

  • Transfer learning based on existing models trained on COCO dataset improves performance

  • Capacities for autolabeling and human-in-the-loop iterative model training

Contents