Release v0.1.7

Added

• OpenVINO plugin examples (#159)
• Dataset validation for classification and detection datasets (#160)
• Arbitrary image extensions in formats (import and export) (#166)
• Ability to set a custom subset name for an imported dataset (#166)
• CLI support for NDR(#178)

Changed

• Common ICDAR format is split into 3 sub-formats (#174)

Fixed

• The ability to work with file names containing Cyrillic and spaces (#148)
• Image reading and saving in ICDAR formats (#174)
• Unnecessary image loading on dataset saving (#176)
• Allowed spaces in ICDAR captions (#182)
• Saving of masks in VOC when masks are not requested (#184)

Release v0.1.6 hotfix

Fixed

• Images with no annotations are exported again in VOC formats (#123)
• Inference result for only one output layer in OpenVINO launcher (#125)

Release v0.1.6

Added

• Icdar13/15 dataset format (#96)
• Laziness, source caching, tracking of changes and partial updating for Dataset (#102)
• Market-1501 dataset format (#108)
• LFW dataset format (#110)
• Support of polygons' and masks' confusion matrices and mismathing classes in diff command (#117)
• Add near duplicate image removal plugin (#113)

Changed

• OpenVINO model launcher is updated for OpenVINO r2021.1 (#100)

Fixed

• High memory consumption and low performance of mask import/export, #53 (#101)
• Masks, covered by class 0 (background) masks, should be exported with holes inside (#104)
• diff command invocation problem with missing class methods (#117)

Release v0.1.5

Added

• WiderFace dataset format (#65, #90)
• Function to transform annotations to labels (#66)
• Dataset splits for classification, detection and re-id tasks (#68, #81)
• VGGFace2 dataset format (#69, #82)
• Unique image count statistic (#87)
• Installation with pip: pip install datumaro

Changed

• Dataset class extended with new operations: save, load, export, import_from, detect, run_model (#71)
• Allowed importing Extractor-only defined formats (in Project.import_from, dataset.import_from and CLI/project import) (#71)
• datum project ... commands replaced with datum ... commands (#84)
• Supported more image formats in ImageNet extractors (#85)
• Allowed adding Importer-defined formats as project sources (source add) (#86)
• Added max search depth in ImageDir format and importers (#86)

Deprecated

• datum project ... CLI context (#84)
• Dataset format Importers will be joined with Extractors in the next release

Fixed

• Allow plugins inherited from Extractor (instead of only SourceExtractor) (#70)
• Windows installation with pip for pycocotools (#73)
• YOLO extractor path matching on Windows (#73)
• Fixed inplace file copying when saving images (#76)
• Fixed labelmap parameter type checking in VOC converter (#76)
• Fixed model copying on addition in CLI (#94)

Release v0.1.4

Added

• CamVid dataset format (#57)
• Ability to install opencv-python-headless dependency with DATUMARO_HEADLESS=1 environment variable instead of opencv-python (#62)

Changed

• Allow empty supercategory in COCO (#54)
• Allow Pascal VOC to search in subdirectories (#50)

Release v0.1.3

Added

• ImageNet and ImageNetTxt dataset formats (#41)

Changed

Deprecated

Removed

Fixed

• Default label-map parameter value for VOC converter (#34)
• Randomness of random split transform (#38)
• Transform.subsets() method (#38)
• Supported unknown image formats in TF Detection API converter (#40)
• Supported empty attribute values in CVAT extractor (#45)

Security

Release v0.1.2

Added

• ByteImage class to represent encoded images in memory and avoid recoding on save (#27)

Changed

• Implementation of format plugins simplified (#22)
• default is now a default subset name, instead of None. The values are interchangeable. (#22)
• Improved performance of transforms (#22)

Removed

• image/depth value from VOC export (#27)

Fixed

• Zero division errors in dataset statistics (#31)

Release v0.1.1

Added

• reindex option in COCO and CVAT format converters (#18)
• Support for relative paths in LabelMe format (#19)
• MOTS png mask format support (#21)

Release v0.1.0

Supported Python versions: 3.6, 3.7, 3.8

Interfaces

• Python API for user code
  • Installation as a package
• A command-line tool for dataset manipulations

Features

• Dataset format support (reading, writing, conversions - any to any)

  • Own format
  • CVAT
  • COCO
  • PASCAL VOC
  • YOLO
  • TF Detection API
  • LabelMe

• Dataset building

  • Composite dataset building
  • Class remapping (project transform remap_labels)
  • Subset splitting (project transform random_split)
  • Dataset filtering (project filter)
  • Dataset merging / updating (project merge)

• Dataset operations

  • Dataset multi-source merging + quality checking + cross-source checking (merge)
  • Annotation transformations (project transform)
  • Dataset info (project info)

• Calculation of statistics for datasets (project stats)

  • Pixel mean, std
  • Object counts, area distribution (detection scenario)
  • Image-Class distribution (classification scenario)
  • Pixel-Class distribution (segmentation scenario)
  • Attributes distribution per label

• Dataset comparison (project diff, project ediff)

  • Annotation-annotation comparison
  • Annotation-inference comparison

• Dataset and model debugging

  • Inference explanation (explain)
    • Black-box approach (RISE paper)
  • Ability to run a model on a dataset, read and write the results
    • OpenVINO
    • Caffe, PyTorch, TensorFlow, MxNet - with Accuracy Checker