Convolutional Prototype Network for Open Set Recognition

Yang, Huinan; Zhang, Xu-Yao; Yin, Fei; Yang, Qing; Liu, Cheng‐Lin

doi:10.1109/tpami.2020.3045079

Cited by 76 publications

(45 citation statements)

References 53 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Wen et al [4] proposed a center loss to learn centers for deep features of each identity and used the centers to reduce intra-class variance. Yang et al [3], [5] proposed the Generalized Convolutional Prototype Learning (GCPL) with a prototype loss, which was used as a regularization to improve the intra-class compactness of the feature representation. For the OSR problem, the prototype helps to reduce intra-class distance of the known classes, but it ignored the potential characteristics of the unknown data, resulting in less effective in reducing the open space risk.…”

Section: Prototype Learningmentioning

confidence: 99%

“…However, the softmax loss only encourages the separability of features, and cannot distinguish the known and unknown classes sufficiently. To make the features more discriminative, several methods [3], [4], [5] utilize the prototype to represent each known class in the embedding feature space and encourage the features of training samples close to the corresponding prototypes. As shown in Fig.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Adversarial Reciprocal Points Learning for Open Set Recognition

Chen,

Peng,

Wang

et al. 2021

Preprint

View full text Add to dashboard Cite

Open set recognition (OSR), aiming to simultaneously classify the seen classes and identify the unseen classes as 'unknown', is essential for reliable machine learning. The key challenge of OSR is how to reduce the empirical classification risk on the labeled known data and the open space risk on the potential unknown data simultaneously. To handle the challenge, we formulate the open space risk problem from the perspective of multi-class integration, and model the unexploited extra-class space with a novel concept Reciprocal Point. Follow this, a novel learning framework, termed Adversarial Reciprocal Point Learning (ARPL), is proposed to minimize the overlap of known distribution and unknown distributions without loss of known classification accuracy. Specifically, each reciprocal point is learned by the extra-class space with the corresponding known category, and the confrontation among multiple known categories are employed to reduce the empirical classification risk. Then, an adversarial margin constraint is proposed to reduce the open space risk by limiting the latent open space constructed by reciprocal points. To further estimate the unknown distribution from open space, an instantiated adversarial enhancement method is designed to generate diverse and confusing training samples, based on the adversarial mechanism between the reciprocal points and known classes. This can effectively enhance the model distinguishability to the unknown classes. Extensive experimental results on various benchmark datasets indicate that the proposed method is significantly superior to other existing approaches and achieves state-of-the-art performance. The code is released on github.com/iCGY96/ARPL.

show abstract

Section: Prototype Learningmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Adversarial Reciprocal Points Learning for Open Set Recognition

Chen,

Peng,

Wang

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Similarly, Jang and Kim [8] introduced a one-vs.-rest network architecture for the DNN output layer and combined class-specific decisions to obtain a robust unknown detection score. Yang et al [33] proposed a convolutional prototype network to learn several prototypes for each known class while leaving room for unknowns in the feature space.…”

Section: Related Workmentioning

confidence: 99%

Synthetic Unknown Class Learning for Learning Unknowns

Jang¹

2021

Preprint

View full text Add to dashboard Cite

This paper addresses the open set recognition (OSR) problem, where the goal is to correctly classify samples of known classes while detecting unknown samples to reject. In the OSR problem, "unknown" is assumed to have infinite possibilities because we have no knowledge about unknowns until they emerge. Intuitively, the more an OSR system explores the possibilities of unknowns, the more likely it is to detect unknowns. Thus, this paper proposes a novel synthetic unknown class learning method that generates unknown-like samples while maintaining diversity between the generated samples and learns these samples. In addition to this unknown sample generation process, knowledge distillation is introduced to provide room for learning synthetic unknowns. By learning the unknown-like samples and known samples in an alternating manner, the proposed method can not only experience diverse synthetic unknowns but also reduce overgeneralization with respect to known classes. Experiments on several benchmark datasets show that the proposed method significantly outperforms other state-of-the-art approaches. It is also shown that realistic unknown digits/letters can be generated and learned via the proposed method after training on the MNIST dataset.

show abstract

“…OSR means not only to classify the known classes seen in the training phase, but also to recognize unknown classes not participated in training as unknown. Many previous works have been proposed such as reconstruction-based methods [19] [20] and prototype network-based methods [21]. Besides, statistic modeling [22] [23] is utilized for new class discovery via out-of-distribution detection.…”

Section: Introductionmentioning

confidence: 99%

Analyzing the Separability of SAR Classification Dataset in Open Set Conditions

Liao

Datcu

Zhang

et al. 2021

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

View full text Add to dashboard Cite

In Synthetic Aperture Radar (SAR) image classification applications, there are three categories of data, including training and benchmark data with fixed classes, as well as actual data in practical applications. A real problem comes that there exist unknown classes not included in training and benchmark data, which is defined as the open set condition. However, little work on recognizing unknown classes and analyzing the separability of SAR datasets has been developed. Motivated by this observation, this paper demonstrates the difficulty of practical classification and analyzes SAR dataset separability in open set conditions. In this paper, SAR separability analyzer (SAR-SA) is proposed to model each known class as a multivariate Gaussian distribution. SAR-SA can classify the known classes and recognize the samples locating in each known distribution with low probabilities as unknown. Besides, SAR dataset-wise separability index (DSI) and class-wise separability index (CSI) are defined to quantify the separability in open set conditions at the dataset level and class level. DSI and CSI are effective indicators of the difficulty of SAR classification datasets. Extensive experimental results demonstrate that the DSI in open set conditions is nearly half of that in supervised conditions. Dataset with low DSI is hard to realize accurate classification in open set conditions. At the class level, even though the SAR image classes are semantically different from each other, there exists more or less overlap between the distributions of supervised known classes and unknown classes. Classes with low CSI are harder to be correctly classified and recognized.

show abstract

Convolutional Prototype Network for Open Set Recognition

Cited by 76 publications

References 53 publications

Adversarial Reciprocal Points Learning for Open Set Recognition

Adversarial Reciprocal Points Learning for Open Set Recognition

Synthetic Unknown Class Learning for Learning Unknowns

Analyzing the Separability of SAR Classification Dataset in Open Set Conditions

Contact Info

Product

Resources

About