Equalization Loss for Long-Tailed Object Recognition

Abstract

Long-tail distribution is widely spread in real-world applications. Due to the extremely small ratio of instances, tail categories often show inferior accuracy. In this paper, we find such performance bottleneck is mainly caused by the imbalanced gradients, which can be categorized into two parts: (1) positive part, deriving from the samples of the same category, and (2) negative part, contributed by other categories. Based on comprehensive experiments, it is also observed that the gradient ratio of accumulated positives to negatives is a good indicator to measure how balanced a category is trained. Inspired by this, we come up with a gradient-driven training mechanism to tackle the long-tail problem: re-balancing the positive/negative gradients dynamically according to current accumulative gradients, with a unified goal of achieving balance gradient ratios. Taking advantage of the simple and flexible gradient mechanism, we introduce a new family of gradient-driven loss functions, namely equalization losses. We conduct extensive experiments on a wide spectrum of visual tasks, including two-stage/single-stage long-tailed object detection (LVIS), long-tailed image classification (ImageNet-LT, Places-LT, iNaturalist), and long-tailed semantic segmentation (ADE20K). Our method consistently outperforms the baseline models, demonstrating the effectiveness and generalization ability of the proposed equalization losses.Codes will be released at https://github.com/ModelTC/United-Perception.

Cross-Modality Interaction Network for Equine Activity Recognition Using Imbalanced Multi-Modal Data

“…In addition, , is the one-hot ground-truth label. However, the models based on softmax CE loss often suffer from inferior classification performance, especially for monitory classes, due to the imbalanced data distribution [ 23 ]. Therefore, we further introduced an effective loss function to supervise the training of CMI-Net and alleviate the class imbalance problem, namely, CB focal loss.…”

“…However, oversampling and undersampling come with high potential risks of overfitting and information loss, respectively [ 21 ]. Reweighting is more flexible and convenient by directly assigning a weight for the loss function per training sample to alleviate the sensitivity of the model to data distribution [ 23 ]. This method is further divided into class-level and sample-level reweighting.…”

“…In our method, softmax cross-entropy (CE) loss was initially used to supervise the training of CMI-Net. However, softmax CE loss suffered from inferior classification performance, especially for monitory classes [ 23 ]. In contrast, CB focal loss, by adding a CB term to focal loss, focuses more on minor-class samples and hard-classified samples and can alleviate the class imbalance problem.…”

Robust Semisupervised Land-Use Classification Using Remote Sensing Data With Weak Labels

“…In 2019, Cao et al alternatively studied the minimum margin per class and designed a label-distribution-aware loss function that encourages a model to have the optimal trade-off between per-class margins [28]. Tan et al proposed equalization loss to tackle the problem of rare long-tailed categories by ignoring the gradients for rare categories [29]. In recent years, all of these methods have become popular reweighting methods.…”

Hierarchical Multi-Label Object Detection Framework for Remote Sensing Images

“…In the final layer of a typical classification module, the representative operation, e.g., SoftMax, is usually used for computing a probability distribution over all classes. In other words, the probabilities across all classes compete against each other and optimization might be biased depending on the category frequency [39]. In the case of HSCN, given the hierarchically annotated dataset where there are hierarchical relationships among classes, we partition the global competition over all classes into the local competitions between sibling classes.…”