CenterNet, rewritten so you might actually finish reading it

What it does

CenterNet-better is a ground-up reimplementation of the “Objects as Points” object detector. The author found the original codebase difficult to parse, so this version is built on a personal dl_lib—itself borrowing structure from detectron2—to prioritize readability. It reports roughly half the training time and 1–2 mAP point improvements over the original repo on standard ResNet backbones.

The interesting bit

The project treats configuration as a first-class object you can inspect at runtime, which turns the usual “where did this hyperparameter come from?” debugging session into a slightly less miserable experience. Training also ends with an auto-generated markdown performance report, so you get a summary without writing your own post-processing glue.

Key highlights

• Claims ~50% faster training and 1–2 mAP points higher than the original CenterNet on COCO (ResNet-18 through ResNet-101).
• Config-as-object design lets you run and inspect configuration values directly instead of hunting through nested dicts or YAML files.
• Vectorized operations replace some slower loops for better throughput.
• Built on a custom dl_lib framework that inherits patterns from detectron2.
• Auto-generates a markdown performance report after training finishes.

Caveats

• ResNet-18 and ResNet-50 training expect over 150 GB of system memory; the README warns you to reduce NUM_WORKER if your hardware falls short.
• DLA, Hourglass backbones, and KeyPoints support are still on the roadmap and not yet implemented.

Verdict

Worth a look if you are studying CenterNet and want a codebase that won’t fight you, or if you need a detectron2-flavored training loop with less boilerplate. Skip it if you need Hourglass or DLA backbones today, or if you are allergic to custom frameworks that are not pure PyTorch.