Most existing methods of semantic segmentation still suffer from two aspects challenges: intra-class inconsistency and inter-class indistinction. To tackle these problems, we propose a Discriminative Feature Network (DFN), which contains sub-networks: Smooth Border Network. Specifically, to handle the problem, specially design with Channel Attention Block global average pooling select more discriminative features. Furthermore, make bilateral features boundary distinguishable deep...

Image dehazing using learning-based methods has achieved state-of-the-art performance in recent years. However, most existing train a model on synthetic hazy images, which are less able to generalize well real images due domain shift. To address this issue, we propose adaptation paradigm, consists of an image translation module and two modules. Specifically, first apply bidirectional network bridge the gap between domains by translating from one another. And then, use before after proposed...

We present an efficient high-resolution network, Lite-HRNet, for human pose estimation. start by simply applying the shuffle block in ShuffleNet to HRNet (high-resolution network), yielding stronger performance over popular lightweight networks, such as MobileNet, ShuffleNet, and Small HRNet. find that heavily-used pointwise (1 × 1) convolutions blocks become computational bottleneck. introduce a unit, conditional channel weighting, replace costly blocks. The complexity of weighting is...

Recent works have widely explored the contextual dependencies to achieve more accurate segmentation results. However, most approaches rarely distinguish different types of dependencies, which may pollute scene understanding. In this work, we directly supervise feature aggregation intra-class and interclass context clearly. Specifically, develop a Context Prior with supervision Affinity Loss. Given an input image corresponding ground truth, Loss constructs ideal affinity map learning Prior....

We present an effective semi-supervised learning algorithm for single image dehazing. The proposed applies a deep Convolutional Neural Network (CNN) containing supervised branch and unsupervised branch. In the branch, neural network is constrained by loss functions, which are mean squared, perceptual, adversarial losses. we exploit properties of clean images via sparsity dark channel gradient priors to constrain network. train on both synthetic data real-world in end-to-end manner. Our...

Temporal action proposal generation aims to estimate temporal intervals of actions in untrimmed videos, which is a challenging yet important task the video understanding field. The proposals generated by current methods still suffer from inaccurate boundaries and inferior confidence used for retrieval owing lack efficient modeling effective boundary context utilization. In this paper, we propose Context Aggregation Network (TCANet) generate high-quality through "local global" aggregation...

Current few-shot action recognition methods reach impressive performance by learning discriminative features for each video via episodic training and designing various temporal alignment strategies. Nevertheless, they are limited in that (a) individual without considering the entire task may lose most relevant information current episode, (b) these strategies fail misaligned instances. To overcome two limitations, we propose a novel Hybrid Relation guided Set Matching (HyRSM) approach...

Current state-of-the-art approaches for few-shot action recognition achieve promising performance by conducting frame-level matching on learned visual features. However, they generally suffer from two limitations: i) the procedure between local frames tends to be inaccurate due lack of guidance force long-range temporal perception; ii) explicit motion learning is usually ignored, leading partial information loss. To address these issues, we develop a Motion-augmented Long-short Contrastive...

Recent real-time semantic segmentation methods usually adopt an additional branch to pursue rich long-range context. However, the incurs undesirable computational overhead and slows inference speed. To eliminate this dilemma, we propose SCTNet, a single CNN with transformer information for segmentation. SCTNet enjoys representations of inference-free while retaining high efficiency lightweight CNN. utilizes as training-only considering its superb ability extract With help proposed...

We present an effective blind image deblurring method based on a data-driven discriminative prior. Our work is motivated by the fact that good prior should favor clear images over blurred ones. In this work, we formulate as binary classifier which can be achieved deep convolutional neural network (CNN). The learned able to distinguish whether input or not. Embedded into maximum posterior (MAP) framework, it helps in various scenarios, including natural, face, text, and low-illumination...

Crowd counting is a concerned yet challenging task in computer vision. The difficulty particularly pronounced by scale variations crowd images. Most state-of-art approaches tackle the multi-scale problem adopting multi-column CNN architectures where different columns are designed with filter sizes to adapt variable pedestrian/object sizes. However, structure bloated and inefficient, it infeasible adopt multiple deep due huge resource cost. We instead propose Scale Pyramid Network (SPN) which...

Person search aims at localizing and identifying a query person from gallery of uncropped scene images. Different re-identification (re-ID), its performance also depends on the localization accuracy pedestrian detector. The state-of-the-art methods train detector individually, detected bounding boxes may be sub-optimal for following re-ID task. To alleviate this issue, we propose driven refinement framework providing refined detection search. Specifically, develop differentiable ROI...

Semantic segmentation requires both rich spatial information and sizeable receptive field. However, modern approaches usually compromise resolution to achieve real-time inference speed, which leads poor performance. In this paper, we address dilemma with a novel Bilateral Segmentation Network (BiSeNet). We first design Spatial Path small stride preserve the generate high-resolution features. Meanwhile, Context fast downsampling strategy is employed obtain sufficient On top of two paths,...

Most recent approaches for online action detection tend to apply Recurrent Neural Network (RNN) capture long-range temporal structure. However, RNN suffers from non-parallelism and gradient vanishing, hence it is hard be optimized. In this paper, we propose a new encoder-decoder framework based on Transformers, named OadTR, tackle these problems. The encoder attached with task token aims the relationships global inter-actions between historical observations. decoder extracts auxiliary...

Most existing methods of semantic segmentation still suffer from two aspects challenges: intra-class inconsistency and inter-class indistinction. To tackle these problems, we propose a Discriminative Feature Network (DFN), which contains sub-networks: Smooth Border Network. Specifically, to handle the problem, specially design with Channel Attention Block global average pooling select more discriminative features. Furthermore, make bilateral features boundary distinguishable deep...

Self-supervised learning presents a remarkable performance to utilize unlabeled data for various video tasks. In this paper, we focus on applying the power of self-supervised methods improve semi-supervised action proposal generation. Particularly, design an effective Semi-supervised Temporal Action Proposal (SSTAP) framework. The SSTAP contains two crucial branches, i.e., temporal-aware branch and relation-aware branch. improves model by introducing temporal perturbations, feature shift...

Existing GAN inversion methods fail to provide latent codes for reliable reconstruction and flexible editing simultaneously. This paper presents a transformer-based image model pretrained StyleGAN which is not only with less distortions, but also of high quality flexibility editing. The proposed employs CNN encoder multi-scale features as keys values. Meanwhile it regards the style code be determined different layers generator queries. It first initializes query tokens learnable parameters...

Recently, many approaches tackle the Unsupervised Domain Adaptive person re-identification (UDA re-ID) problem through pseudo-label-based contrastive learning. During training, a uni-centroid representation is obtained by simply averaging all instance features from cluster with same pseudo label. However, may contain images different identities (label noises) due to imperfect clustering results, which makes inappropriate. In this paper, we present novel Multi-Centroid Memory (MCM) adaptively...

Improving boundary segmentation results has recently attracted increasing attention in the field of semantic segmentation. Since existing popular methods usually exploit long-range context, cues are obscure feature space, leading to poor results. In this paper, we propose a novel conditional loss (CBL) for improve performance boundaries. The CBL creates unique optimization goal each pixel, conditioned on its surrounding neighbors. is easy yet effective. contrast, most previous boundary-aware...