One-Shot Image Classification by Learning to Restore Prototypes

Xue, Wanqi; Wang, Wei

doi:10.1609/aaai.v34i04.6130

Cited by 39 publications

(18 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…However, this is sometimes not enough when we have only 1 picture for a rare class. In this case, [13] introduces architecture to press the image with noise closer to the center of the class.…”

Section: Classification Of Rare Classesmentioning

confidence: 99%

Rare Traffic Signs Recognition

Badmaev

Shakhuro

Konushin

2021

Proceedings of the 31th International Conference on Computer Graphics and Vision. Volume 2

View full text Add to dashboard Cite

Recognition of road signs is an important part of the control systems of autonomous vehicles and driver assistance systems. Modern recognition methods based on neural networks require large well-labeled datasets. Marking up data is quite expensive, but it is even more difficult to mark up rare classes of objects. To solve this problem in this article, we use synthetic data. We improve the marking of the Russian traffic signs dataset (RTSD) in semi-automatic mode adding 9 thousand new road signs. We perform an experimental evaluation of the currently best classifiers and detectors in the task of recognizing road signs. To improve the performance of classification, we use stochastic weight averaging (SWA) and contrastive loss. The use of modern methods allows us to train a high-quality neural network on synthetic data, which was previously impossible, and significantly improves the metrics of recognition of both rare and frequent road signs.

show abstract

Section: Classification Of Rare Classesmentioning

confidence: 99%

Rare Traffic Signs Recognition

Badmaev

Shakhuro

Konushin

2021

Proceedings of the 31th International Conference on Computer Graphics and Vision. Volume 2

View full text Add to dashboard Cite

show abstract

“…Another key issue with smoke detection is that performance is considerably degraded in forest environments that are different from the training set [16]. The success of these deep methods can be partially attributed to a large number of annotated data [17]. However, forest fire smoke images or videos are very hard to capture in real life.…”

Section: Introductionmentioning

confidence: 99%

Adversarial Fusion Network for Forest Fire Smoke Detection

Zhang

Zhu

et al. 2022

Forests

View full text Add to dashboard Cite

Recent advances suggest that deep learning has been widely used to detect smoke for early forest fire warnings. Despite its remarkable success, this approach has a number of problems in real life application. Deep neural networks only learn deep and abstract representations, while ignoring shallow and detailed representations. In addition, previous models have been trained on source domains but have generalized weakly on unseen domains. To cope with these problems, in this paper, we propose an adversarial fusion network (AFN), including a feature fusion network and an adversarial feature-adaptation network for forest fire smoke detection. Specifically, the feature fusion network is able to learn more discriminative representations by fusing abstract and detailed features. Meanwhile, the adversarial feature adaptation network is employed to improve the generalization ability and transfer gains of the AFN. Comprehensive experiments on two self-built forest fire smoke datasets, and three publicly available smoke datasets, validate that our method significantly improves the performance and generalization of smoke detection, particularly the accuracy of the detection of small amounts of smoke.

show abstract

“…The observations motivate us to learn to estimate more accurate and representative prototypes instead of to fine-tune the feature extractor. Recently, [30] also attempts to estimate more representative prototypes by learning a mapping function from noisy samples to its ground-truth center. However, learning to recover representative prototypes from noisy samples without any priors is very challenging.…”

Section: Introductionmentioning

confidence: 99%

“…In this paper, we propose to address the FSL problem with a novel prototype completion framework. Different from [30], we leverage the primitive knowledge (i.e., classlevel attribute or part annotations) [24], e.g., whether a class object should have ears, legs, or eyes, to enable a metalearner to learn to complete prototypes. Specifically, our proposed framework first introduces the primitive knowledge and extracts attribute features from the base class samples as priors.…”

Section: Introductionmentioning

confidence: 99%

Prototype Completion with Primitive Knowledge for Few-Shot Learning

Zhang

et al. 2021

2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

105

View full text Add to dashboard Cite

Few-shot learning is a challenging task, which aims to learn a classifier for novel classes with few labeled samples. Previous studies mainly focus on two-phase meta-learning methods. Recently, researchers find that introducing an extra pre-training phase can significantly improve the performance. The key idea is to learn a feature extractor with pretraining and then fine-tune it through the nearest centroid based meta-learning. However, results show that the finetuning step makes very marginal improvements. We thus argue that the current meta-learning scheme does not fully explore the power of the pre-training. The reason roots in the fact that in the pre-trained feature space, the base classes already form compact clusters while novel classes spread as groups with large variances. In this case, finetuning the feature extractor is less meaningful than estimating more representative prototypes. However, making such an estimation from few labeled samples is challenging because they may miss representative attribute features. In this paper, we propose a novel prototype completion based meta-learning framework. The framework first introduces primitive knowledge (i.e., class-level attribute or part annotations) and extracts representative attribute features as priors. A prototype completion network is then designed to learn to complement the missing attribute features with the priors. Finally, we develop a Gaussian based prototype fusion strategy to combine the mean-based and the complemented prototypes, which can effectively exploit the unlabeled samples. Extensive experimental results on three real-world data sets demonstrate that our method: (i) can obtain more accurate prototypes; (ii) outperforms state-ofthe-art techniques by 2% ∼ 9% on classification accuracy.

show abstract

One-Shot Image Classification by Learning to Restore Prototypes

Cited by 39 publications

References 9 publications

Rare Traffic Signs Recognition

Rare Traffic Signs Recognition

Adversarial Fusion Network for Forest Fire Smoke Detection

Prototype Completion with Primitive Knowledge for Few-Shot Learning

Contact Info

Product

Resources

About