CenterNet-better

Why we included this project

The original CenterNet code is powerful but notoriously hard to follow, and this reimplementation exists to fix that. It keeps the Objects as Points idea intact while making the codebase easier to read, and the author reports it needs roughly half the training time for 1 to 2 points higher mAP. Configuration becomes a debuggable object, several operations are vectorized for speed, and the repo evaluates your model and writes a markdown report automatically after training. If you already work in PyTorch and want a cleaner base for keypoint-based detection, this is a much more approachable starting point than the upstream repo, and it doubles as a readable reference for understanding how CenterNet works internally or building a custom detector on a ResNet backbone.