Bounding boxes that actually fit

What it does

MMRotate is an OpenMMLab toolbox for rotated object detection built on PyTorch. It provides a modular training and inference framework where components can be swapped to assemble new models without rewriting boilerplate. The project ships with a model zoo covering roughly twenty algorithms, from Rotated RetinaNet to Oriented R-CNN and the newer RTMDet family.

The interesting bit

Instead of forcing every object into an axis-aligned rectangle, the framework supports three mainstream angle representations natively, letting researchers pick the geometric flavor that matches their target domain—useful when horizontal boxes waste pixels on background noise.

Key highlights

• Modular design decomposes detectors into pluggable components (backbone, neck, head, etc.)
• Three angle representations available to match different paper implementations
• Model zoo includes Rotated Faster R-CNN, Oriented R-CNN, GWD, KLD, and RTMDet
• RTMDet variant reports 78.9 AP (single-scale) / 81.3 AP (multi-scale) on DOTA at 121 FPS on an RTX 3090
• Apache 2.0 license

Verdict

Worth a look if you need oriented bounding boxes and want to benchmark against a wide zoo of published methods. Skip it if you are after a minimal, dependency-free detector.