Half-Real Half-Fake Distillation for Class-Incremental Semantic Segmentation

Huang, Zilong; Hao, Wentian; Wang, Xinggang; Tao, Mingyuan; Huang, Jianqiang; Liu, Wenyu; Hua, Xian-Sheng

doi:10.48550/arxiv.2104.00875

Cited by 2 publications

(3 citation statements)

References 48 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Under the Overlapped setting, we can see that the performance of our method consistently outperforms that of other methods by a sizable margin on all evaluated VOC benchmarks (i.e., 19-1, 15-5 and 15-1). On VOC-19-1, the forgetting of old classes (1)(2)(3)(4)(5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19) is reduced by 1.91% while performance on new classes is greatly improved by 18.25%. On the most challenging benchmark VOC-15-1, it is worth noting that the performance of our method on all the seen classes outperforms its closest contender PLOP [12] by around 7.48%.…”

Section: Comparison To State-of-the-art Methodsmentioning

confidence: 99%

“…The latter is more challenging because the task identity is unavailable at inference time. Recently, continual learning has been also explored on several other computer vision tasks, e.g., incremental object detection [23], incremental video classification [48], incremental instance segmentation [15], continual semantic segmentation [3,12,14,18,31,33,39,43,45,52]. Our work focuses on the CSS problem which can be considered as the classincremental learning scenario on semantic segmentation.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

RBC: Rectifying the Biased Context in Continual Semantic Segmentation

Zhao¹,

Yang²,

Fu³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recent years have witnessed a great development of Convolutional Neural Networks in semantic segmentation, where all classes of training images are simultaneously available. In practice, new images are usually made available in a consecutive manner, leading to a problem called Continual Semantic Segmentation (CSS). Typically, CSS faces the forgetting problem since previous training images are unavailable, and the semantic shift problem of the background class. Considering the semantic segmentation as a context-dependent pixel-level classification task, we explore CSS from a new perspective of context analysis in this paper. We observe that the context of old-class pixels in the new images is much more biased on new classes than that in the old images, which can sharply aggravate the old-class forgetting and new-class overfitting. To tackle the obstacle, we propose a biased-context-rectified CSS framework with a context-rectified image-duplet learning scheme and a biased-context-insensitive consistency loss. Furthermore, we propose an adaptive re-weighting class-balanced learning strategy for the biased class distribution. Our approach outperforms state-of-the-art methods by a large margin in existing CSS scenarios.

show abstract

Section: Comparison To State-of-the-art Methodsmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

RBC: Rectifying the Biased Context in Continual Semantic Segmentation

Zhao¹,

Yang²,

Fu³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…For instance, RECALL-GAN [28] utilizes a pre-trained generative model to retrieve old knowledge. Huang et al [152] propose to use a pretrained image-generative model to invert the trained segmentation network to synthesize input images from random noise. Besides, pre-trained models can be used as an auxiliary task to boost the CSS task.…”

Section: Regularization-based Mannermentioning

confidence: 99%

Ecosystem Health of the Beiyun River Basin (Beijing, China) as Evaluated by the Analytic Hierarchy Process Method

Yuan

Zhao

et al. 2021

Preprint

View full text Add to dashboard Cite

Ecosystem services provided by river ecosystems rely on healthy ecosystem structure and ecological processes. The Beijing-Tianjin-Hebei urban is a typical water-deficient area. As an important part of the urban-rural integration construction, evaluating the health status of the Beiyun River Basin and discovering the weak links in the water environment is the basis for improving the health of the basin. In this study, Analytic Hierarchy Process was used to establish an evaluation index system for the Beiyun River Basin from 5 aspects including water quality, biology, habitat, hydrology, and social functions, and to assign weights to the index layer. The evaluation results showed that the health evaluation results of the Beiyun River Basin in 2019 are “sub-healthy”, and the overall health status is gradually worsening from northwest to southeast. In the middle reaches of the region, the evaluation result is "healthy", followed by the upstream, and the downstream is the worst. The results showed that areas with less human interference are in better health. The factors that affect the overall health evaluation status in the basin are the level of nutrition, biodiversity, and vegetation coverage. For the comprehensive management of the Beiyun River, the improvement of water quality and habitat ecological restoration is the key to the health of the upstream ecosystem health status. The improvement of the health status of the downstream should focus on equal emphasis on water quality and quantity, restoration of biodiversity, and improvement of the quality of the riparian ecological environment.

show abstract

Half-Real Half-Fake Distillation for Class-Incremental Semantic Segmentation

Cited by 2 publications

References 48 publications

RBC: Rectifying the Biased Context in Continual Semantic Segmentation

RBC: Rectifying the Biased Context in Continual Semantic Segmentation

Ecosystem Health of the Beiyun River Basin (Beijing, China) as Evaluated by the Analytic Hierarchy Process Method

Contact Info

Product

Resources

About