Long-tail recognition tackles the natural non-uniformly distributed data in real-world scenarios. While modern classiﬁers perform well on populated classes, its performance degrades signiﬁcantly on tail classes. Humans, however, are less aﬀected by this since, when confronted with uncertain examples, they simply opt to provide coarser predictions. Motivated by this, a deep realistic taxonomic classiﬁer (Deep-RTC) is proposed as a new solution to the long-tail problem, combining realism with hierarchical predictions. The model has the option to reject classifying samples at diﬀerent levels of the taxonomy, once it cannot guarantee the desired performance. Deep-RTC is implemented with a stochastic tree sampling during training to simulate all possible classiﬁcation conditions at ﬁner or coarser levels and a rejection mechanism at inference time. Experiments on the long-tailed version of four datasets, CIFAR100, AWA2, Imagenet, and iNaturalist, demonstrate that the proposed approach preserves more information on all classes with diﬀerent popularity levels. Deep-RTC also outperforms the state-of-the-art methods in longtailed recognition, hierarchical classiﬁcation, and learning with rejection literature using the proposed correctly predicted bits (CPB) metric.
Architecture: Schematic representation of our network.
Training, evaluation and deployment code available on GitHub.
This work was partially funded by NSF awards IIS-1637941, IIS-1924937, and NVIDIA GPU donations.