Adaptive Logit Adjustment Loss for Long-Tailed Visual Recognition

Zhao, Yan; Chen, Weicong; Tan, Xu; Huang, Kai; Zhu, Jihong

doi:10.1609/aaai.v36i3.20258

Cited by 23 publications

(17 citation statements)

References 34 publications

(67 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We randomly select ten images per class from training data to construct metadata. Several classical and advanced robust loss functions and data augmentation approaches that are mainly designed for long-tailed classifications are compared, including Classbalanced CE loss [31], Class-balanced Focal loss, LDAM [23], LDAM-DRW [23], ISDA [7], LA [24], ALA [33], LPL [32], MixUp [35], and RISDA [9]. Besides, De-confound-TDE [34], which uses causal intervention in training and counterfactual reasoning in inference, is also involved in our comparison.…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

A General Multiple Data Augmentation Based Framework for Training Deep Neural Networks

Sun

Qin

2022

2022 International Joint Conference on Neural Networks (IJCNN)

View full text Add to dashboard Cite

Machine-learning models are prone to capturing the spurious correlations between non-causal attributes and classes, with counterfactual data augmentation being a promising direction for breaking these spurious associations. However, explicitly generating counterfactual data is challenging, with the training efficiency declining. Therefore, this study proposes an implicit counterfactual data augmentation (ICDA) method to remove spurious correlations and make stable predictions. Specifically, first, a novel sample-wise augmentation strategy is developed that generates semantically and counterfactually meaningful deep features with distinct augmentation strength for each sample. Second, we derive an easy-to-compute surrogate loss on the augmented feature set when the number of augmented samples becomes infinite. Third, two concrete schemes are proposed, including direct quantification and meta-learning, to derive the key parameters for the robust loss. In addition, ICDA is explained from a regularization aspect, with extensive experiments indicating that our method consistently improves the generalization performance of popular depth networks on multiple typical learning scenarios that require out-of-distribution generalization.

show abstract

Section: Methodsmentioning

confidence: 99%

“…Methods designed for long-tailed classification including Class-balanced CE loss [31], OLTR [40], LDAM [23], LDAM-DRW [23], LA [24], ALA [33], RISDA [9], Meta-classweight [36], and MetaSAug [8] are compared. Results.…”

Section: Methodsmentioning

confidence: 99%

A General Multiple Data Augmentation Based Framework for Training Deep Neural Networks

Sun

Qin

2022

2022 International Joint Conference on Neural Networks (IJCNN)

View full text Add to dashboard Cite

show abstract

“…To cope with this problem, GCL [14] introduces Gaussian clouds into the logit adjustment process, and adaptively sets the cloud size according to the sample size of each class. Recently, Zhao et al [46] reveal that previous logit adjustment techniques primarily focus on the sample quantity of each class, while ignoring the difficulty of samples. Thus, Zhao et al [46] propose to prevent the over-optimization on easy samples of tail classes, while highlighting the training on difficult samples of head classes.…”

Section: Balanced Lossmentioning

confidence: 99%

“…Recently, Zhao et al [46] reveal that previous logit adjustment techniques primarily focus on the sample quantity of each class, while ignoring the difficulty of samples. Thus, Zhao et al [46] propose to prevent the over-optimization on easy samples of tail classes, while highlighting the training on difficult samples of head classes.…”

Section: Balanced Lossmentioning

confidence: 99%

Revisiting Long-tailed Image Classification: Survey and Benchmarks with New Evaluation Metrics

Fang¹,

Zhang²,

Wen³

et al. 2023

Preprint

View full text Add to dashboard Cite

Recently, long-tailed image classification harvests lots of research attention, since the data distribution is long-tailed in many real-world situations. Piles of algorithms are devised to address the data imbalance problem by biasing the training process towards less frequent classes. However, they usually evaluate the performance on a balanced testing set or multiple independent testing sets having distinct distributions with the training data. Considering the testing data may have arbitrary distributions, existing evaluation strategies are unable to reflect the actual classification performance objectively. We set up novel evaluation benchmarks based on a series of testing sets with evolving distributions. A corpus of metrics are designed for measuring the accuracy, robustness, and bounds of algorithms for learning with long-tailed distribution. Based on our benchmarks, we re-evaluate the performance of existing methods on CI-FAR10 and CIFAR100 datasets, which is valuable for guiding the selection of data rebalancing techniques. We also revisit existing methods and categorize them into four types including data balancing, feature balancing, loss balancing, and prediction balancing, according the focused procedure during the training pipeline.

show abstract

“…Also, it is feasible for various places to look similar, particularly within similar environments, an error called perceptual aliasing [4]. Primarily applied for computer vision (CV) tasks, Convolution Neural Network (CNN) based model has been pioneered in the VPR fields during the past few years, obtaining remarkable achievements on different data sets [5].…”

Section: Introductionmentioning

confidence: 99%

Seagull Optimization With Deep Learning Driven Condition Invariant Visual Place Recognition Model

Sasikumar¹,

Sathiamoorthy²

2022

INDJCSE

View full text Add to dashboard Cite

Visual place recognition (VPR) is most important topic in the computer vision (CV) and robotics community, whereas the task is for efficiently and accurately recognize the place of a provided query image. VPR still remains an open problem because of the several manners whereas the presence of real-world locations can change. As condition-invariant and viewpoint-invariant features were important features to long-term robust visual place-detection, the accomplishment of deep learning (DL) approaches from the CV domain triggered a kind of primary studies as its efficacy for VPR utilizing generic features in networks which are trained for another kind of detection tasks. In this aspect, this article develops a Seagull Optimization with Deep Learning Driven Condition Invariant Visual Place Recognition (SGODL-CIVPR) model. The presented SGODL-CIVPR model majorly involves the recognition of live places with respect to reference images. To accomplish this, the SGODL-CIVPR technique utilizes deep convolutional neural network based EfficientNet model to produce features, and the hyperparameter tuning process takes place using the SGO algorithm. Besides, convolutional wavelet neural network (CWNN) with deep convolutional generative adversarial network (DCGAN) is used to map features to low dimensional space for improving invariant properties and decrease dimensions concurrently. The performance validation of the SGODL-CIVPR model is tested using different datasets and the results demonstrate better performance over other methods.

show abstract

Adaptive Logit Adjustment Loss for Long-Tailed Visual Recognition

Cited by 23 publications

References 34 publications

A General Multiple Data Augmentation Based Framework for Training Deep Neural Networks

A General Multiple Data Augmentation Based Framework for Training Deep Neural Networks

Revisiting Long-tailed Image Classification: Survey and Benchmarks with New Evaluation Metrics

Seagull Optimization With Deep Learning Driven Condition Invariant Visual Place Recognition Model

Contact Info

Product

Resources

About