Long-Tailed Classification Based on Coarse-Grained Leading Forest and Multi-Center Loss

TL;DR

This work tackles the generalized long-tailed problem by addressing both class-wise and attribute-wise imbalances. It introduces Cognisance, a framework that combines an unsupervised Coarse-Grained Leading Forest (CLF) for environment construction with a Multi-Center Loss (MCL) for invariant feature learning, designed to plug into existing LT methods. A noise-selection mechanism and a Cognisance^+ extension with a Multi-Center Triplet Loss (MCTL) further enhance robustness to mislabeled tail data. Experiments on ImageNet-GLT and MSCOCO-GLT demonstrate state-of-the-art performance and improved generalization across CLT and GLT settings, including noisy data scenarios, highlighting practical impact for real-world imbalanced datasets.

Abstract

Long-tailed (LT) classification is an unavoidable and challenging problem in the real world. Most existing long-tailed classification methods focus only on solving the class-wise imbalance while ignoring the attribute-wise imbalance. The deviation of a classification model is caused by both class-wise and attribute-wise imbalance. Due to the fact that attributes are implicit in most datasets and the combination of attributes is complex, attribute-wise imbalance is more difficult to handle. For this purpose, we proposed a novel long-tailed classification framework, aiming to build a multi-granularity classification model by means of invariant feature learning. This method first unsupervisedly constructs Coarse-Grained forest (CLF) to better characterize the distribution of attributes within a class. Depending on the distribution of attributes, one can customize suitable sampling strategies to construct different imbalanced datasets. We then introduce multi-center loss (MCL) that aims to gradually eliminate confusing attributes during feature learning process. The proposed framework does not necessarily couple to a specific LT classification model structure and can be integrated with any existing LT method as an independent component. Extensive experiments show that our approach achieves state-of-the-art performance on both existing benchmarks ImageNet-GLT and MSCOCO-GLT and can improve the performance of existing LT methods. Our codes are available on GitHub: \url{https://github.com/jinyery/cognisance}

Figures (10)

• Figure 1: Inter-class long-tailed distribution and attribute-wise long-tailed distribution.
• Figure 2: Spurious correlation of the "White" attribute with the "Swan" category. Firstly, black swans are more likely to be misclassified than white swans in the category "Swan", even though they both belong to the same category. Secondly, the attribute "White" may be falsely correlated with "Swan", so when "White" appears in images of "Cock", there is a high risk that the "cock" will be misclassified as a "Swan".
• Figure 3: A leading tree constructed from digit '2' in MNIST dataset (taken from our previous work xu2021lapoleaf), in which each path can reflect an implicit attribute within the same class. Note that this figure is used only to explain our intuition and motivation, we did not evaluate our method on MNIST dataset.
• Figure 4: The left is an example of CLF constructed for category "sand", while the right is an example of attribute splitting using CLF. Each path from the root to the leaf node can be considered as an (implicit) attribute, and the samples within the coarse-grained nodes are very similar that require an appropriate reduction in sampling weights. In addition, the samples within the red and pink boxes demonstrate the potential of CLF for noisy recognition.
• Figure 5: Overall framework diagram. Different environments have different sampling strategies, where $q_{cls}$ and $q_{attr}$ are balancing factors for class-wise sampling and attribute-wise sampling, respectively.