Research map / Label-efficient and robust learning / Recognition and long-tail

Feature Fusion from Head to Tail for Long-Tailed Visual Recognition

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Decoupled and calibrated long-tail training · 52 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes head-to-tail fusion (H2T), a plug-and-play classifier-stage method that swaps random channels of tail-class feature maps with head-class feature maps, improving long-tailed recognition on CIFAR100-LT, ImageNet-LT, iNaturalist 2018 and Places-LT.

Problem

Long-tailed visual recognition suffers from biased decision boundaries because tail classes have sparse semantic regions, leading many tail samples to be misclassified as head classes.

Approach

Augment tail classes by grafting a fraction of feature-map channels from head-class samples onto tail-class features (using balanced and instance-wise sampling branches), without labels mixing, and fine-tune only the classifier in a second stage, so the backbone is unchanged.

Key results

• On CIFAR100-LT (imbalance ratio 100), RIDE+H2T reaches 51.38% vs 50.37% for RIDE, and GCL+H2T reaches 48.88%.
• On iNaturalist 2018, DR+H2T achieves 71.81% top-1, and MisLAS+H2T reaches 72.05%, the best single-model result reported in its table.
• On CIFAR10-LT with imbalance ratio 200, DR+H2T improves over DR with CE loss by more than 5 points (78.59%).

Measured effects reported in the paper

Surprising findings

• Replacing all feature maps with another class still yields 43.57% on CIFAR100-LT, above CE loss by 4.02%.
• H2T gives smaller gains for GCL and for cosine classifiers than for linear classifiers.

Limitations

• Improves tail accuracy only by adjusting the decision boundary; the embedding distribution is unchanged and head accuracy drops as the fusion ratio grows.
• Random channel selection is used; the authors note a principled selection strategy remains future work.
• Gains are smaller for cosine-classifier and GCL setups.

Experimental setup

Datasets: CIFAR100-LT ImageNet-LT iNaturalist 2018 Places-LT CIFAR10-LT

Benchmarks: CIFAR100-LT ImageNet-LT iNaturalist 2018 Places-LT CIFAR10-LT

Metrics: top-1 accuracy head/medium/tail accuracy

Baselines: LDAM-DRW DR MisLAS GCL MBJ CMO ABL BBN RIDE ACE ResLT

Builds on: decoupling representation (cRT) MisLAS GCL CutMix Manifold MixUp

Closest papers by meaning

• Bag of Tricks for Long-Tailed Visual Recognition with Deep Convolutional Neural Networks (AAAI 2021 · similarity 0.91)
• Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · similarity 0.90)
• Deep Representation Learning on Long-Tailed Data: A Learnable Embedding Augmentation Perspective (CVPR 2020 · similarity 0.90)
• Long-Tailed Recognition via Information-Preservable Two-Stage Learning (NeurIPS 2025 · similarity 0.89)
• Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.88)
• DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.88)