Changelog.

What every version of fotonet added, changed and fixed. Newest first.

  1. 1.2

    Trainer 1.2

    Latest

    The same fotonetn, retrained with a new trainer. 32.1 mAP50-95, up from 29.9.

    The release page

    Changed

    • fotonetn is retrained with the 1.2 trainer: 32.1 mAP50-95 on COCO val2017, up from 29.9, and 47.5 mAP50, up from 42.9. Same model, same size, same speed.
    • Fewer duplicate boxes: at confidence 0.3, duplicates fell from 16.1% to 6.1%.
    • More precise at the default confidence of 0.25: precision rose from 49% to 70%, with 62% fewer false positives. Recall there fell from 49% to 44%; F1, which weighs both, rose from 49.0 to 54.1.

    Added

    • A warmup, stable and decay learning-rate schedule, with the image size stepping up during training.
    • A forecast of the final mAP while a run trains.
    • Live settings: change how often a run validates and previews without stopping it.
    • Guard rails that measure duplicates, precision and recall during training.

  2. 1.1.1

    Loading fixes

    Model names load the right weights, and fine-tuning starts from them.

    Fixed

    • fotonete loads from the v1 release and fotonetn from v1.1.
    • Fine-tuning starts from the weights you loaded instead of a fresh model.

  3. 1.1

    UltraDense

    A new family of NMS-free detectors, built from scratch. fotonetn comes first.

    Added

    • UltraDense, a new family of detectors in four sizes with one API: fotonetn, fotonets, fotonetm and fotonetl. Every one is NMS-free and trained from scratch.
    • fotonetn weights: 29.9 mAP50-95 on COCO val2017, from 2.85M parameters and 6.95 GFLOPs.
    • The fotonets, fotonetm and fotonetl architectures. Their weights follow.
    • Video tracking: model.track() keeps an ID on every object across frames.

    Changed

    • fotonete, the v1 model, stays supported.

  4. 1.0.1

    Load models by name

    The released checkpoint loads by name.

    Changed

    • Fotonet("fotonete") loads the verified release checkpoint. You no longer download it and pass a path.

  5. 1.0.0

    fotonet 1.0

    The first stable release, with fotonete, the v1 model.

    Added

    • fotonete, the v1 model, as the supported model.
    • Predict on images, folders and arrays, and on OpenCV frames with predict_bgr.
    • Train on your own data from a YAML dataset, and resume an interrupted run.
    • Export to ONNX.
    • The fotonet command line for predicting, training and exporting.

  6. Previews

    Alpha and beta releases on PyPI before 1.0.

  7. 0.8 beta

    Beta

    Preview

    One shared model design, chainable box transforms, and the fotonet name.

    Added

    • Chainable box transforms: set an anchor, expand by pixels, clamp to the image and crop, in one line.
    • P2 variants with an extra high-resolution level, for small-object experiments.

    Changed

    • Every public model shares one backbone, neck and head implementation.
    • The project is called fotonet, in lowercase.

  8. 0.2 to 0.3 alpha

    Five alphas in three weeks

    Preview

    Ending with automatic export checks.

    Added

    • ONNX and TorchScript exports are checked against PyTorch automatically.

    Changed

    • Named sizes such as fotonetn are architectures until you load a checkpoint. Accuracy is published only with a released checkpoint and the command that measured it.

  9. 0.1 alpha

    First alpha

    Preview

    The first public preview, for experiments and feedback.

    Added

    • NMS-free object detection: the network learns one-to-one assignment, so each object gets one box.
    • Optional one-to-many supervision while training.
    • Plain Python results, and transform helpers for crops, anchors, moving and resizing boxes.
    • Export to ONNX, TorchScript and TensorRT. CoreML export was included but not yet verified.
    • fotonet-n, an alpha nano checkpoint: 22.68% mAP50-95 on COCO from 2.75M parameters.