Convolutional Fine-Grained Classification with Self-Supervised Target Relation Regularization

TL;DR

This paper proposes a novel self-supervised target relation regularization method using dynamic target relation graphs for fine-grained classification, improving robustness and performance on imbalanced and sparse data.

Contribution

Introduces dynamic target relation graphs (DTRG), a self-generated structural regularization scheme that models inter-class relations for better fine-grained classification.

Findings

Achieves state-of-the-art results on fine-grained benchmarks.

Enhances robustness against data sparsity and imbalance.

Effectively models inter-class dependencies through DTRG.

Abstract

Fine-grained visual classification can be addressed by deep representation learning under supervision of manually pre-defined targets (e.g., one-hot or the Hadamard codes). Such target coding schemes are less flexible to model inter-class correlation and are sensitive to sparse and imbalanced data distribution as well. In light of this, this paper introduces a novel target coding scheme -- dynamic target relation graphs (DTRG), which, as an auxiliary feature regularization, is a self-generated structural output to be mapped from input images. Specifically, online computation of class-level feature centers is designed to generate cross-category distance in the representation space, which can thus be depicted by a dynamic graph in a non-parametric manner. Explicitly minimizing intra-class feature variations anchored on those class-level centers can encourage learning of discriminative features. Moreover, owing to exploiting inter-class dependency, the proposed target graphs can alleviate data sparsity and imbalanceness in representation learning. Inspired by recent success of the mixup style data augmentation, this paper introduces randomness into soft construction of dynamic target relation graphs to further explore relation diversity of target classes. Experimental results can demonstrate the effectiveness of our method on a number of diverse benchmarks of multiple visual classification tasks, especially achieving the state-of-the-art performance on popular fine-grained object benchmarks and superior robustness against sparse and imbalanced data. Source codes are made publicly available at https://github.com/AkonLau/DTRG.