Search

Learning disentangled representations from data that reveal inherent semantic information is crucial for the interpretability of generative models such as Variational Auto-Encoders (VAEs) and Generative Adversarial Networks (GANs). However, finding a balance between disentanglement and generation is challenging. The traditional disentangled model, β -VAE, often fails to reconstruct clear images. To address this issue, we propose a novel contrastive learning method called β -CLVAE that enhances the generation ability and improves the disentanglement of β -VAE. In our method, we use a Recurrent Neural Network (RNN) to generate robust anchor samples that are fed, along with the reconstructions from β -VAE, to a contrastive learning framework. We also introduce a pixel-level Gram matrix that imposes constraints on the input and output images to improve the quality of the decoder's reconstruction. By using contrastive learning, β -CLVAE can learn valuable semantic information that benefits disentangled and generative representations. Our experiments demonstrate the advantages of our model in terms of qualitative comparison of disentangled features, quantitative comparison of disentanglement scores, and the accuracy of downstream classification tasks. Overall, β -CLVAE is an excellent semantic disentangled representation model that is beneficial for downstream tasks. [ABSTRACT FROM AUTHOR]

Weakly supervised temporal action localization (W-TAL) is designed to detect and classify all the action instances in an untrimmed video with only video-level labels. Due to the lack of frame-level annotations, the correlations learning between action snippets and the separation between action and background are the two key issues for accurate action localization. To mine the intrinsic correlations of space and time embodied in the occurrences of action in a video and identify the action and background in the snippets, a novel method based on spatial–temporal correlations learning and action-background jointed attention for W-TAL is proposed. In this method, the graph convolution and 1-D temporal convolution networks are constructed to learn the spatial and temporal features of the video, respectively, then fused to learn a fruitful spatial–temporal correlative feature map. This ensures more completed features representation for action localization. Next, different from the other methods, action-background jointed attention mechanism is presented to explicitly modelled background as well as action in a three-branch classification network. This classification network can distinguish action and background and realize the separation of action and background better, so as to promote more accurate action localization. Experiments conducted on Thumos14 and ActivityNet1.3 show that our method outperforms state-of-the-art methods, especially at some high t-IoU thresholds, which further validates the effectiveness of our method. [ABSTRACT FROM AUTHOR]

During the era of Industry 4.0, the industrial robot monitoring process is getting success and popularity day by day. It also plays a vital role in the enhancement of robot vision algorithms. This paper proposed a model Deep Three Dimensional Spearman Correlation Analysis (D3D-SCA) to address nonlinear dimensionality reduction in robot vision for three-dimensional data. Dealing with three-dimensional multimedia datasets using traditional algorithms, to date, researchers have been facing limitations and challenges because mostly sub-space learning algorithms and their developments cannot perform satisfactorily in most of the time with linear and non-linear data dependency. The proposed model directly finds the relations between two sets of three-dimensional data without reshaping the data into 2D-matrices or vectors and dramatically reduces the dimensional reduction and computational algorithm complexity. The proposed model extracts deep information and translates it into a decision. To do so, three components are employed in the proposed model: customized deep learning model Inception_V3 for deep feature mapping, three-dimensional spearman correlation analysis for comparing pairwise deep features without a singular matrix and spatial dilemma problem, and the customized Xception classifier with automatic online updating ability and adjustable neural architecture for low latency models. The motivation of the proposed model is to advance the scalability of existing industrial robot vision applications which based on recognition, detection and re-identification approaches. Extensive findings on industrial datasets named "3D Objects on turntable and Caltech 101" demonstrate the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) have obtained high accuracy results for pedestrian re-identification in the past few years. There is always a trade-off between high accuracy and computational time in CNNs. Training CNN is always very difficult as it may take a long time to produce high accuracy results. To overcome this limitation, a novel method parallel stochastic gradient descent (PSGD) is proposed to train a five-hierarchical parallel CNNs that is designed according to pedestrian attributes. Moreover, the momentum correction and adaptive adjustment of learning rate are applied during training process and the time interval for updating parameters is inspected during optimization of parameters selection. The results of this paper prove the effectiveness of proposed PSGD that successfully decreases the training process by five times and surpasses the state-of-the-art methods of pedestrian re-identification in terms of both accuracy and time. The minimum reported running time of the proposed method is 8.7 s which is minimum among all other state-of-the-art methods. These promising results show the efficiency and performance of the proposed model. [ABSTRACT FROM AUTHOR]

Generative Adversarial Networks (GANs) are most popular generative frameworks that have achieved compelling performance. They follow an adversarial approach where two deep models generator and discriminator compete with each other. They have been used for many applications especially for image synthesis because of their capability to generate high quality images. In past few years, different variants of GAN have proposed and they produced high quality results for image generation. This paper conducts an analysis of working and architecture of GAN and its popular variants for image generation in detail. In addition, we summarize and compare these models according to different parameters such as architecture, training method, learning type, benefits and performance metrics. Finally, we apply all these methods on a benchmark MNIST dataset, which contains handwritten digits and compare qualitative and quantitative results. The evaluation is based on quality of generated images, classification accuracy, discriminator loss, generator loss and computational time of these models. The aim of this study is to provide a comprehensive information about GAN and its various models in the field of image synthesis. Our main contribution in this work is critical comparison of popular GAN variants for image generation on MNIST dataset. Moreover, this paper gives insights regarding existing limitations and challenges faced by GAN and discusses associated future research work. [ABSTRACT FROM AUTHOR]

Searching for specific persons from surveillance videos captured by different cameras, known as person re-identification, is a key yet under-addressed challenge. Difficulties arise from the large variations of human appearance in different poses, and from the different camera views that may be involved, making low-level descriptor representation unreliable. In this paper, we propose a novel Sparse Representations based Distributed Attribute Learning Model (SRDAL) to encode targets into semantic topics. Compared to other models such as ELF, our model performs best by imposing semantic restrictions onto the generation of human specific attributes and employing Deep Convolutional Neural Network to generate features without supervision for attributes learning model. Experimental results show that our method achieves state-of-the-art performance. [ABSTRACT FROM AUTHOR]