OpenLongTailRecognition-OLTR

Why we included this project

Long-tail data is the norm in many real-world recognition tasks, and OLTR is the authors' own PyTorch implementation of their CVPR 2019 oral method for handling it. The repo comes with training configs and pretrained weights for the ImageNet-LT and Places-LT benchmarks, so you can reproduce the reported numbers and compare OLTR against simpler techniques like reweighting or resampling. That makes it a practical reference for teams that need to improve recall on rare classes, whether those are uncommon species, unusual defect types, or niche product categories, without giving up accuracy on frequent ones. The code also fixes the false positive metric calculation and separates stage one and stage two training cleanly, so it is easy to adapt if you bring your own data.