Equalization Loss for Long-Tailed Object Recognition

By Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, Junjie Yan. 2020

556 citations recorded by the source corpus.

Object recognition techniques using convolutional neural networks (CNN) have achieved great success. However, state-of-the-art object detection methods still perform poorly on large vocabulary and long-tailed datasets, e.g. LVIS. In this work, we analyze this problem from a novel perspective: each positive sample of one category can be seen as a negative sample for other categories, making the tail categories receive more discouraging gradients. Based on it, we propose a simple but effective loss, named equalization loss, to tackle the problem of long-tailed rare categories by simply ignoring those gradients for rare categories. The equalization loss protects the learning of rare categories from being at a disadvantage during the network parameter updating. Thus the model is capable of learning better discriminative features for objects of rare classes. Without any bells and whistles, our method achieves AP gains of 4.1% and 4.8% for the rare and common categories on the challenging LVIS benchmark, compared to the Mask R-CNN baseline. With the utilization of the effective equalization loss, we finally won the 1st place in the LVIS Challenge 2019. Code has been made available at: https://github.com/tztztztztz/eql.detectron2

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Optimization & Training

Citations in the corpus

Reviving Iterative Training with Mask Guidance for Interactive Segmentation (2021, background): “These datasets stimulated the development of state-of-the-art instance segmentation algorithms [42, 43, 44, 45], providing a large variety of labeled classes with a sufficient number of examples.”
MDETR - Modulated Detection for End-to-End Multi-Modal Understanding (2021, background): “Federated datasets often pose problems to standard detectors, and require developing specific loss functions [48].”
1st Place Solution of LVIS Challenge 2020: A Good Box is not a Guarantee of a Good Mask (2020, methodology): “At the representation learning stage, we use some techniques like EQL [15], repeat factor re-sampling [5], data augmentation, self-training to learn generalized representation.”
Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect (2020, unknown): “In this paper, we propose a causal framework that not only fundamentally explains the previous methods [15–17, 9, 11, 10], but also provides a principled solution to further improve long-tailed classification.”
Distribution Alignment: A Unified Framework for Long-tail Visual Recognition (2021, unknown): “Class-level methods typically re-weight the standard loss with category-specific coefficients correlated with the sample distributions [18, 9, 3, 21, 20, 34].”
Learning to Segment the Tail (2020, background): “’s work [19] to further improve our performances.”
CReST: A Class-Rebalancing Self-Training Framework for Imbalanced Semi-Supervised Learning (2021, background): “Prominent works include re-sampling [7, 3, 4, 14] and re-weighting [21, 9, 6, 41] which re-balance the contribution of each class, while others focus on re-weighting each instance [27, 38, 35, 19].”
Detecting Twenty-thousand Classes using Image-level Supervision (2022, background): “Equalization losses [49, 50] and SeeSaw loss [58] reweights the per-class loss by balancing the gradients [49] or number of samples [58].”
Balanced Activation for Long-tailed Visual Recognition (2020, unknown): “Equalization Loss [9]: The solution of the winner of LVIS Challenge 2019.”
Simple Multi-dataset Detection (2021, background): “OpenImages models treat rare classes differently [38] and model the hierarchy of classes in the loss [26].”
Long-tail learning via logit adjustment (2020, background): “…of work involve post-hoc normalisation of the classification weights (Zhang et al., 2019; Kim & Kim, 2019; Kang et al., 2020; Ye et al., 2020), and modification of the underlying loss to account for varying class penalties (Zhang et al., 2017; Cui et al., 2019; Cao et al., 2019; Tan et al., 2020).”
Contrastive Learning based Hybrid Networks for Long-Tailed Image Classification (2021, methodology): “Typical methods include data re-sampling [1, 8, 35, 4, 28], loss re-weighting [26, 7, 30, 33], margin modification [3], and data augmentation [19, 6, 29].”
Probabilistic two-stage detection (2021, methodology): “To push the state-of-the-art of object detection, we further switch to a stronger backbone Res2Net (Gao et al., 2019a) with BiFPN (Tan et al., 2020b), a larger input resolution ( 1280 × 1280 in training and 1560 × 1560 in testing) with heavy crop augmentation (ratio 0.1 to 2) (Tan et al., 2020b),…”
Seesaw Loss for Long-Tailed Instance Segmentation (2020, methodology): “This differs signiﬁcantly to previous solutions that rely either on static group split [26] or loss reweighting with constant values [2, 7, 44].”
Balanced Knowledge Distillation for Long-tailed Learning (2021, methodology): “In the literature, one of the most common practices to facilitate learning for tail classes is to rebalance the class distribution, either by re-sampling the examples [9, 4, 6, 26] or re-weighting the classification loss [18, 23, 7, 46].”