Major update featuring significant corrections to exposure compensation algorithm, improved quality scoring model, many UI improvements, bug fixes, compatibility improvements, security improvements, and more.
Major Changes
- New Exposure compensation solver algorithm ensures that the bird in the image is always properly exposed. This improves quality ranking accuracy, species detection accuracy, and makes it easier to review your photos!
- New setting to tweak exposure compensation solver performance to your needs.
- Substantial user experience improvements:
- New "Quick Copy" buttons let you directly copy bird crops or whole image exports to your clipboard.
- New adjustable divider lets you resize the bird crop to full image preview space.
- New "Suggested Species" tags make it easier to add tags based on machine learning model outputs.
- Improved split scene behavior (use Shift+Click to split scenes).
- "Only manually reviewed scenes" toggle now reflects culling decisions, scenes with reviewed tags, and renamed scenes.
- Improved multi-bird handling:
- Kestrel now detects, analyzes, and saves results for up to 5 birds within the same image.
- Browse through detected birds using up/down arrow keys; hit "Enter" to assign a primary bird.
- Model & Performance Updates:
- Upgraded quality classification model for substantially improved performance.
- Improved performance on scenes with many subjects via model parameter tweaks.
- Security improvements (reduced HTTP endpoint exposure, settings schema validation).
- Improved crash handling/recovery with automatic queue restoration.
Minor Changes
- Fixed bugs on Canon EOS R5 Mk II by upgrading rawpy dependency.
- Upgraded tensorflow, torch, and other dependencies to latest versions.
- Improved auto-save performance and consistency.
- Fixed bugs related to loading/unloading folders during analysis.
- Fixed bug where JPG files would not be co-moved via culling assistant.
- Fixed bugs with scene navigation.
Known Issues
- Analysis time may be slightly slower due to the new exposure algorithm; use "Normal" or "Lenient" profiles to reduce overhead.
- Default star rating normalization algorithm is being refined.
Coming soon
- Website will be updated to reflect latest analysis framework and new features.
Feedback? Please visit projectkestrel.org/contact or use the in-app feedback feature!