Exploiting multi-scale representations is critical to improve edge detection for objects at different scales. To extract edges dramatically scales, we propose a Bi-Directional Cascade Network (BDCN) structure, where an individual layer supervised by labeled its specific scale, rather than directly applying the same supervision all CNN outputs. Furthermore, enrich learned BDCN, introduce Scale Enhancement Module (SEM) which utilizes dilated convolution generate features, instead of using...

Exploiting multi-scale representations is critical to improve edge detection for objects at different scales. To extract edges dramatically scales, we propose a bi-directional cascade network (BDCN) architecture, where an individual layer supervised by labeled its specific scale, rather than directly applying the same supervision layers. Furthermore, enrich learned each of BDCN, introduce scale enhancement module (SEM), which utilizes dilated convolution generate features, instead using...

Recently unsupervised domain adaptation for the semantic segmentation task has become more and popular due to high-cost of pixel-level annotation on real-world images. However, most methods are only restricted single-source-single-target pair, can not be directly extended multiple target domains. In this work, we propose a collaborative learning framework achieve multi-target adaptation. An expert model is first trained each source-target pair further encouraged collaborate with other...

Multi-source unsupervised domain adaptation (MSDA) aims at adapting models trained on multiple labeled source domains to an unlabeled target domain. In this paper, we propose a novel multi-source framework based collaborative learning for semantic segmentation. Firstly, simple image translation method is introduced align the pixel value distribution reduce gap between and some extent. Then, fully exploit essential information across domains, without seeing any data from addition, similar...

Semi-supervised learning is a useful tool for image segmentation, mainly due to its ability in extracting knowledge from unlabeled data assist labeled data. This paper focuses on popular pipeline known as self-learning, where we point out weakness named lazy mimicking that refers the inertia model retains prediction itself and thus resists updates. To alleviate this issue, propose Asynchronous Teacher-Student Optimization (ATSO) algorithm (i) breaks up continual teacher student (ii)...

Data-driven based approaches, in spite of great success many tasks, have poor generalization when applied to unseen image domains, and require expensive cost annotation especially for dense pixel prediction tasks such as semantic segmentation. Recently, both unsupervised domain adaptation (UDA) from large amounts synthetic data semi-supervised learning (SSL) with small set labeled been studied alleviate this issue. However, there is still a gap on performance compared their supervised...

Universal domain adaptation (UniDA) aims to transfer the knowledge learned from a label-rich source label-scarce target without any constraints on label space. However, shift and category make UniDA extremely challenging, which mainly lies in how recognize both shared "known" samples private "unknown" samples. Previous works rarely explore intrinsic geometrical relationship between two domains, they manually set threshold for overconfident closed-world classifier reject Therefore, this...

Recently, contrastive learning has largely advanced the progress of unsupervised visual representation learning. Pre-trained on ImageNet, some self-supervised algorithms reported higher transfer performance compared to fully-supervised methods, seeming deliver message that human labels hardly contribute transferrable features. In this paper, we defend usefulness semantic but point out and methods are pursuing different kinds To alleviate issue, present a new algorithm named Supervised...

Exploiting multi-scale representations is critical to improve edge detection for objects at different scales. To extract edges dramatically scales, we propose a Bi-Directional Cascade Network (BDCN) structure, where an individual layer supervised by labeled its specific scale, rather than directly applying the same supervision all CNN outputs. Furthermore, enrich learned BDCN, introduce Scale Enhancement Module (SEM) which utilizes dilated convolution generate features, instead of using...

Multi-source unsupervised domain adaptation~(MSDA) aims at adapting models trained on multiple labeled source domains to an unlabeled target domain. In this paper, we propose a novel multi-source adaptation framework based collaborative learning for semantic segmentation. Firstly, simple image translation method is introduced align the pixel value distribution reduce gap between and some extent. Then, fully exploit essential information across domains, without seeing any data from addition,...

Recently unsupervised domain adaptation for the semantic segmentation task has become more and popular due to high-cost of pixel-level annotation on real-world images. However, most methods are only restricted single-source-single-target pair, can not be directly extended multiple target domains. In this work, we propose a collaborative learning framework achieve multi-target adaptation. An expert model is first trained each source-target pair further encouraged collaborate with other...

Data-driven based approaches, in spite of great success many tasks, have poor generalization when applied to unseen image domains, and require expensive cost annotation especially for dense pixel prediction tasks such as semantic segmentation. Recently, both unsupervised domain adaptation (UDA) from large amounts synthetic data semi-supervised learning (SSL) with small set labeled been studied alleviate this issue. However, there is still a gap on performance compared their supervised...

With its outstanding performance and tracking speed, discriminative correlation filters (DCF) have gained much attention in visual object tracking, where time-consuming operations can be efficiently computed utilizing the discrete Fourier transform (DFT) with symmetric properties. Nevertheless, inherent issues of boundary effects filter degradation, as well occlusion background clutter, degrade performance. In this work, we proposed an augmented memory joint aberrance repressed (AMRCF) for...

The discriminative correlation filter (DCF)-based tracking method has shown good accuracy and efficiency in visual tracking. However, the periodic assumption of sample space causes unwanted boundary effects, restricting tracker's ability to distinguish between target background. Additionally, real environment, interference factors such as occlusion, background clutter, illumination changes cause response aberration and, thus, failure. To address these issues, this work proposed a novel named...

Weight pruning is a well-known technique used for network compression. In contrast to filter pruning, weight produces higher compression ratios as it more fine-grained. However, individual weights results in broken kernels, which cannot be directly accelerated on general platforms, leading hardware compatibility issues. To address this issue, we propose Shift Pruning (SP), novel method that compatible with platforms. SP converts spatial convolutions into regular 1 X and shift operations, are...

Image understanding is a fundamental task for many multimedia and computer vision applications, such as self-driving, retrieval, augmented reality, etc. In this paper, we demonstrate that edge detection could aid image tasks semantic segmentation, optical flow estimation, object proposal generation. Based on our recent research efforts detection, develop robust efficient Edge-Aided imaGe undERstanding system named EAGER. EAGER built compact module, which constructed with bi-directional...

In medical image analysis, semi-supervised learning is an effective method to extract knowledge from a small amount of labeled data and large unlabeled data. This paper focuses on popular pipeline known as self learning, points out weakness named lazy that refers the difficulty for model learn pseudo labels generated by itself. To alleviate this issue, we propose ATSO, asynchronous version teacher-student optimization. ATSO partitions into two subsets alternately uses one subset fine-tune...

While discriminative correlation filter (DCF) has attracted much attention due to its excellent computational efficiency and robustness, factors like occlusion, motion blur, deformation background interference cause tracking failure. To address these issues, this work proposes improved background-aware (BACF) that utilize colour features as a complement of Histogram Oriented Gradient (HOG) improve the representation target, implementing adaptive feature fusion in response layer using Peak...