v0.11.0

[0.11.0] - 2026-07-16

Added

• File menu: Reset Settings to Defaults resets Model and Crop settings. Recent folders and window layout remain untouched.

Changed

• The cropped and sorted images all land in a processed subdirectory of the directory that is currently being viewed instead of one named after the confidence and model. Images are no longer split into cropped/non-cropped subfolders. Filenames get _crop/_ncrop suffixes instead.
• The detections CSV and settings.json filenames now encode the confidence and model, so multiple runs' summaries can coexist.
• The detections CSV's last column is renamed "Subdirectory" -> "cropped" (yes/no).
• Cropping/sorting multiple images no longer auto-opens the file manager. This prevents focus stealing and the output folder can still be accessed using the link of the toast message or the Open output folder menu entry.

v0.10.2

[0.10.2] - 2026-07-09

Added

• Images already analyzed (and instantly available from the cache) are now shown in dark green in the image list.

Changed

• Faster image display and detection preprocessing as the 8-bit conversion now goes directly from the source color layout and skips redundant full-frame copies.
• Improved chaching: stepping back one image during forward browsing (or vice versa) no longer discards the cached images ahead, so quickly bouncing back and forth stays instant.
• Recent folder paths on the welcome screen now show ~ instead of the full home directory path.
• [Dev] Add logging module.
• [Dev] Add a Detection NamedTuple replacing raw result tuples, and simplify the Detector constructor.
• [Dev] Add a simple benchmark (poe bench) for the 8-bit conversion pipeline.

Fixed

• Fix reading, writing, and drag-and-drop on Windows for deeply nested folders or long file names, matching the write-side fix in 0.9.1.
• Fix the detection model being loaded twice at startup.

v0.10.1

[0.10.1] - 2026-07-08

Fixed

• Fix a hang that stalled cropping/export, caused by missing toast notification assets.

v0.10.0

[0.10.0] - 2026-07-08

Added

• The welcome screen now lists recently opened folders to provide a quick way to reopen them.
• New "Actions > Open output folder" menu item reveals the most recent export directory.
• Show toast notifications when:
  • crop and sort actions finish (plus a "Show in file manager" link)
  • background model downloads finish or fail
  • filenames are copied to the clipboard
  • image processing fails
• "Clear Image List" is now available in the image list context menu.

Fixed

• Removing a large selection of images from the list is now instant instead of freezing the UI.
• Fix a Qt SSL warning that appeared ~30s after a model download completed.
• Detect macOS AppleDouble sidecar files (e.g. "._IMG_1234.HIF") and exclude them from import.

Changed

• The progress dialog for batch runs now shows immediately and guards against starting a second run.
• Single-image actions don't open the file manager automatically anymore.
• The "recent folder no longer exists" warning is now a toast instead of a modal dialog.

v0.9.2

[0.9.2] - 2026-07-06

Changed

• Detectorist prefetches and processes the next 3 images in the current direction. Cached images & detections are displayed almost instantly so the app feels snappier.
• The "Loading image..." placeholder only appears when loading actually takes noticeable time.
• Batch runs load the next image in the background while the current one is processed, cutting batch time (up to 2x for HIF files).
• A batch run no longer aborts when one image fails to load. The image is skipped and recorded as "load-error" in detections.csv.
• Rename the "Crop to largest area" option to "Crop to union of detected objects". Previously saved crop settings still load.
• The displayed image now reflects the "Auto correct camera exposure bias" setting, matching what gets exported.
• [Dev] Extract crop planning from the main window into a crop_planner module.
• [Dev] Extract batch processing (crop & export, sort by class) into a batch_run module.
• [Dev] Concentrate the exposure correction and EXIF update logic in the ImageObject base class.
• [Dev] Upgrade dependencies.
• [Dev] Cleanup CI and collapse build jobs into a single matrix job. Update release version handling.

Fixed

• Exposure-corrected JPEG crops now reset the EXIF exposure bias to 0, as HEIF and other formats already did. Correcting an already corrected crop no longer doubles the adjustment.

v0.9.1

[0.9.1] - 2026-06-25

Changed

• [Dev] Make release script handle HTTPS git remotes.
• [Dev] Upgrade dependencies.

Fixed

• Fix export/cropping on Windows for deeply nested folders or long file names. Output paths over the 260 char MAX_PATH limit now use Windows extended-length paths.

v0.9.0

[0.9.0] - 2026-03-10

Added

• Add support for instance segmentation. Three new models are included:
  • Fish Segmentation
  • Apoidea Segmentation
  • Generic Instance Segmentation
• The model manager now detects when an older version of a model is installed and marks it as outdated, with an option to download the newer version.

Changed

• Model names in the combo box no longer include the release date. It is still shown in the model manager.
• [Dev] Refactor model management code.
• [Dev] Rename generated Qt UI files to the ui_* convention.
• [Dev] QRC omit per-file timestamps for reproducible builds.
• [Dev] Upgrade dependencies.

v0.8.2

[0.8.2] - 2026-02-21

Changed

• Inference speed-up due to upgraded dependencies.

Fixed

• Fix macOS x86_64 build.

v0.8.1

[0.8.1] - 2026-02-19

Added

• Support to browse and download detection models from the project page.
• A "Generic Object Detection" model that detects 80 everyday object classes (Person, Bicycle, Car, etc.) based on RF-DETR.
• Added a filter combo box to filter displayed detections by object class.
• Local-only models (on disk but not available on the remote) also appear in the model dialog.
  transformer.
• The remote model manifest is cached to disk after the first successful fetch so that human-readable model names are available immediately on the next launch, even before the manifest is re-fetched.

Changed

• The application prompts to download models at first start and doesn't ship them with the release binary anymore.
• The model selector shows the human-readable model name (e.g. "Fish Detection Model") instead of the raw filename.
• The confidence slider now filters bounding boxes instantly without running inference again.
• Models are distributed as gzip-compressed .onnx.gz files, reducing download size.

v0.8.0

[0.8.0] - 2026-02-16

Added

• Support for DETR (DEtection TRansformer) models for better object detection and faster inference.

Changed

• Upgrade detection models:
  • apoidea-detect-transformer-2026-02-16 replaces bee-detect-2025-09-10.
  • fish-detect-transformer-2026-02-15 replaces fish-detect-2025-09-11.
  • Thanks to the High Performance and Cloud Computing Group at the Zentrum für Datenverarbeitung of the University of Tübingen for providing the computing resources to train train our models on the bwForCluster BinAC 2.