Reproduce “Feature Generation for Long-tail Classification” on OpenResearch

About this paper

Rahul Vigneswaran, Marc T. Law, Vineeth N. Balasubramanian, Makarand Tapaswi

TailCalibration introduces a method to generate synthetic feature representations for under-represented classes in long-tail datasets, aiming to balance data for a decoupled classifier. The dynamic variant, TailCalibX, establishes new state-of-the-art performance on CIFAR-100-LT and mini-ImageNet-LT, demonstrating accuracy improvements of over 20% for tail classes on mini-ImageNet-LT.

Key insights

• Generating features directly in the latent space effectively addresses data scarcity and diversity for tail classes without the computational complexity and challenges of generating raw images.
• TailCalibration is orthogonal and complementary to existing advanced long-tail techniques, demonstrated by its ability to further boost performance when combined with state-of-the-art methods like Class-Balanced Distillation (CBD).
• Dynamic feature generation (TailCalibX) during classifier training provides consistent incremental performance improvements, indicating the benefits of continuous exposure to varied synthetic data.

Summary by alphaXiv.