luminoth

Why we included this project

Luminoth is worth a look if you want a tidy, self-contained way to train and run object detectors without assembling the pieces yourself. It bundles Faster R-CNN and SSD behind a single `lumi` command, and you can adapt a dataset, train locally or on Google Cloud ML Engine, and export checkpoints trained on COCO and Pascal. The project was officially abandoned in 2020, with the authors recommending Detectron2 for anything modern, so treat it as a study aid rather than a dependency. That said, the Sonnet-based architecture and the separation between training, evaluation, and detection code make it easy to read and borrow from. It will be most useful to people learning how backbones, region proposals, and anchor boxes fit together, or maintaining an older TensorFlow 1.x stack.