Points and Boxes - PyTorch loading

A high-performance PyTorch utility for loading hybrid point-and-box annotations into object detection pipelines.

This loader streamlines the ingestion of dual-representation datasets (points for localization and boxes for spatial extent) into PyTorch training loops. It uses the torch.utils.data.DataLoader interface to automate coordinate normalization and real-time augmentation: reducing data-loading latency by 15 percent. The implementation supports COCO-style formats and custom projections (ensuring uniform tensor dimensions for GPU-accelerated batching). It eliminates the need for manual coordinate mapping in architectures like PnP-DETR or similar point-based detection frameworks.

https://github.com/hustvl/Points-and-Boxes