v1.0.0

Glasses Detector v1.0.0

About

First stable¹ release for Glasses Detector. Three tasks are supported for processing images with glasses: classification, detection, and segmentation. This includes glasses types and parts, as well as their shadows and eye-area. For all the available features and examples, please check documentation and notebook.

Note

Pre-trained weights for size large are not available as it seems most medium models perform better (less overfit due to tasks being binary). More experiments need to be conducted before large weights can be released in v1.1.0.

What's New

• Features
  • Glasses and eye-area detection
  • More categories for each task
  • Better model performance
  • Simpler CLI
• Architecture
  • New package structure
  • New model architectures
  • Simpler customization process (better inheritance structure)
• Data
  • More datasets
  • Simpler data processing scripts
  • Kaggle notebooks for sub-tasks
• Documentation
  • Full documentation of package-specific code
  • Features & examples
  • Fancier documentation page style

Other

As before, although the code for training (e.g., datasets, metrics) is part of PIP package, the actual data processing, training, and model analysis scripts are only available when fully cloning the repository. Please check Data and Running sections for details on how to train/test your own models.

1. Kind of stable since large weights are missing (technically not but they perform worse than medium) ↩