Long-tail recognition that admits when it meets a stranger

What it does

Open Long-Tailed Recognition (OLTR) is the authors’ PyTorch reimplementation of their CVPR 2019 Oral paper. It trains vision models on heavily imbalanced datasets—where a few classes dominate and most are starved for examples—while also preparing them for open-world test sets that contain categories never encountered during training. The goal is to correctly classify the familiar many-shot and few-shot classes, and honestly flag the rest as unknown rather than forcing a wrong label.

The interesting bit

Most projects treat long-tail classification and open-set detection as separate afflictions; this one treats them as the same headache. The README is essentially a reproduction guide, so the actual machinery—modulatedatt, meta-embedding, and the two-stage training pipeline—remains inside the paper, but the provided configs and weights let you run the full ImageNet-LT and Places-LT benchmarks out of the box.

Key highlights

• Reimplementation of a CVPR 2019 Oral presentation with published model weights and benchmark tables.
• Two-stage training: stage 1 learns representations, stage 2 learns a meta-embedding module.
• Evaluated on standard long-tail benchmarks ImageNet-LT and Places-LT, reporting many-shot, medium-shot, few-shot, and F-measure metrics.
• Stage 1 for Places-LT fine-tunes a ResNet-152 from Caffe pre-trained weights; stage 2 drops back to single-GPU training.
• Reproduced results are updated post-release after bug fixes in attention modules and F-measure calculation.

Caveats

• Multi-GPU training is fragile outside of Places-LT stage 1, and stage 2 explicitly requires a single GPU.
• The README is heavy on directory structure and command flags but light on model intuition; expect to read the paper to understand the architecture.
• PyTorch >= 0.4.1 and Caffe pre-trained weights place the codebase in an earlier era of the framework.

Verdict

Worth a look if you are reproducing long-tail or open-set vision baselines from 2019. Skip it if you need a modern, self-contained library with clean multi-GPU support and extensive documentation.