• Journal of Information Processing Systems
• /
• v.18 no.6
• /
• pp.794-802
• /
• 2022

The recovery of reasonable depth information from different scenes is a popular topic in the field of computer vision. For generating depth maps with better details, we present an efficacious monocular depth prediction framework with coordinate attention and feature fusion. Specifically, the proposed framework contains attention, multi-scale and feature fusion modules. The attention module improves features based on coordinate attention to enhance the predicted effect, whereas the multi-scale module integrates useful low- and high-level contextual features with higher resolution. Moreover, we developed a feature fusion module to combine the heterogeneous features to generate high-quality depth outputs. We also designed a hybrid loss function that measures prediction errors from the perspective of depth and scale-invariant gradients, which contribute to preserving rich details. We conducted the experiments on public RGBD datasets, and the evaluation results show that the proposed scheme can considerably enhance the accuracy of depth prediction, achieving 0.051 for log10 and 0.992 for δ<1.25³ on the NYUv2 dataset.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.2
• /
• pp.79-86
• /
• 2024

In this paper, a novel visual tracking method which can utilize the context of object regions is presented. Conventional methods have the inherent problem of treating all candidate regions independently, where the tracker could not successfully discriminate regions with similar appearances. This was due to lack of contextual modeling in a given scene, where all candidate object regions should be taken into consideration when choosing a single region. The goal of the proposed method is to encourage feature exchange between candidate regions to improve the discriminability between similar regions. It improves upon conventional methods that only consider a single region, and is implemented by employing the MLP-Mixer model for enhanced feature exchange between regions. By implementing channel-wise, inter-region interaction operation between candidate features, contextual information of regions can be embedded into the individual feature representations. To evaluate the performance of the proposed tracker, the large-scale LaSOT dataset is used, and the experimental results show a competitive AUC performance of 0.560 while running at a real-time speed of 65 fps.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.5
• /
• pp.1396-1412
• /
• 2023

Conversation modeling is an important and challenging task in the field of natural language processing because it is a key component promoting the development of automated humanmachine conversation. Most recent research concerning conversation modeling focuses only on the current utterance (considered as the current question) to generate a response, and thus fails to capture the conversation's logic from its beginning. Some studies concatenate the current question with previous conversation sentences and use it as input for response generation. Another approach is to use an encoder to store all previous utterances. Each time a new question is encountered, the encoder is updated and used to generate the response. Our approach in this paper differs from previous studies in that we explicitly separate the encoding of the question from the encoding of its context. This results in different encoding models for the question and the context, capturing the specificity of each. In this way, we have access to the entire context when generating the response. To this end, we propose a deep neural network-based model, called the Context Model, to encode previous utterances' information and combine it with the current question. This approach satisfies the need for context information while keeping the different roles of the current question and its context separate while generating a response. We investigate two approaches for representing the context: Long short-term memory and Convolutional neural network. Experiments show that our Context Model outperforms a baseline model on both ConvAI2 Dataset and a collected dataset of conversational English.

• Biomolecules & Therapeutics
• /
• v.30 no.1
• /
• pp.98-116
• /
• 2022

The small GTPase RhoA has been studied extensively for its role in actin dynamics. In this study, multiple bioinformatics tools were applied cooperatively to the microarray dataset GSE64714 to explore previously unidentified functions of RhoA. Comparative gene expression analysis revealed 545 differentially expressed genes in RhoA-null cells versus controls. Gene set enrichment analysis (GSEA) was conducted with three gene set collections: (1) the hallmark, (2) the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway, and (3) the Gene Ontology Biological Process. GSEA results showed that RhoA is related strongly to diverse pathways: cell cycle/growth, DNA repair, metabolism, keratinization, response to fungus, and vesicular transport. These functions were verified by heatmap analysis, KEGG pathway diagramming, and direct acyclic graphing. The use of multiple gene set collections restricted the leakage of information extracted. However, gene sets from individual collections are heterogenous in gene element composition, number, and the contextual meaning embraced in names. Indeed, there was a limit to deriving functions with high accuracy and reliability simply from gene set names. The comparison of multiple gene set collections showed that although the gene sets had similar names, the gene elements were extremely heterogeneous. Thus, the type of collection chosen and the analytical context influence the interpretation of GSEA results. Nonetheless, the analyses of multiple collections made it possible to derive robust and consistent function identifications. This study confirmed several well-described roles of RhoA and revealed less explored functions, suggesting future research directions.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.8
• /
• pp.319-330
• /
• 2021

Recently, in the field of open domain natural language question answering, multi-task, multi-hop question answering has been studied extensively. In this paper, we propose a novel deep neural network model using hierarchical graphs to answer effectively such multi-task, multi-hop questions. The proposed model extracts different levels of contextual information from multiple paragraphs using hierarchical graphs and graph neural networks, and then utilize them to predict answer type, supporting sentences and answer spans simultaneously. Conducting experiments with the HotpotQA benchmark dataset, we show high performance and positive effects of the proposed model.

• Journal of Korea Multimedia Society
• /
• v.25 no.5
• /
• pp.685-694
• /
• 2022

The goal of few-shot semantic segmentation is to build a network that quickly adapts to novel classes with extreme data shortage regimes. Most existing few-shot segmentation methods leverage single or multiple prototypes from extracted support features. Although there have been promising results for natural images, these methods are not directly applicable to the aerial image domain. A key factor in few-shot segmentation on aerial images is to effectively exploit information that is robust against extreme changes in background and object scales. In this paper, we propose a Mask-Guided Attention module to extract more comprehensive support features for few-shot segmentation in aerial images. Taking advantage of the support ground-truth masks, the area correlated to the foreground object is highlighted and enables the support encoder to extract comprehensive support features with contextual information. To facilitate reproducible studies of the task of few-shot semantic segmentation in aerial images, we further present the few-shot segmentation benchmark iSAID-, which is constructed from a large-scale iSAID dataset. Extensive experimental results including comparisons with the state-of-the-art methods and ablation studies demonstrate the effectiveness of the proposed method.

• ETRI Journal
• /
• v.44 no.6
• /
• pp.1034-1046
• /
• 2022

The learning-based multiview stereo (MVS) methods for three-dimensional (3D) reconstruction generally use 3D volumes for depth inference. The quality of the reconstructed depth maps and the corresponding point clouds is directly influenced by the spatial resolution of the 3D volume. Consequently, these methods produce point clouds with sparse local regions because of the lack of the memory required to encode a high volume of information. Here, we apply the atrous spatial pyramid pooling (ASPP) module in MVS methods to obtain dense feature maps with multiscale, long-range, contextual information using high receptive fields. For a given 3D volume with the same spatial resolution as that in the MVS methods, the dense feature maps from the ASPP module encoded with superior information can produce dense point clouds without a high memory footprint. Furthermore, we propose a 3D loss for training the MVS networks, which improves the predicted depth values by 24.44%. The ASPP module provides state-of-the-art qualitative results by constructing relatively dense point clouds, which improves the DTU MVS dataset benchmarks by 2.25% compared with those achieved in the previous MVS methods.

• Journal of Intelligence and Information Systems
• /
• v.29 no.3
• /
• pp.419-437
• /
• 2023

As the number of thin filers in Korea surpasses 12 million, there is a growing interest in enhancing the accuracy of assessing their credit default risk to generate additional revenue. Specifically, researchers are actively pursuing the development of default prediction models using machine learning and deep learning algorithms, in contrast to traditional statistical default prediction methods, which struggle to capture nonlinearity. Among these efforts, Graph Neural Network (GNN) architecture is noteworthy for predicting default in situations with limited data on thin filers. This is due to their ability to incorporate network information between borrowers alongside conventional credit-related data. However, prior research employing graph neural networks has faced limitations in effectively handling diverse categorical variables present in credit information. In this study, we introduce the Transformer embedded Graph Convolutional Network (TeGCN), which aims to address these limitations and enable effective default prediction for thin filers. TeGCN combines the TabTransformer, capable of extracting contextual information from categorical variables, with the Graph Convolutional Network, which captures network information between borrowers. Our TeGCN model surpasses the baseline model's performance across both the general borrower dataset and the thin filer dataset. Specially, our model performs outstanding results in thin filer default prediction. This study achieves high default prediction accuracy by a model structure tailored to characteristics of credit information containing numerous categorical variables, especially in the context of thin filers with limited data. Our study can contribute to resolving the financial exclusion issues faced by thin filers and facilitate additional revenue within the financial industry.

• KIPS Transactions on Software and Data Engineering
• /
• v.9 no.3
• /
• pp.91-100
• /
• 2020

Unlike the existing Visual Question Answering(VQA) problems, the new Visual Commonsense Reasoning(VCR) problems require deep common sense reasoning for answering questions: recognizing specific relationship between two objects in the image, presenting the rationale of the answer. In this paper, we propose a novel deep neural network model, KG_VCR, for VCR problems. In addition to make use of visual relations and contextual information between objects extracted from input data (images, natural language questions, and response lists), the KG_VCR also utilizes commonsense knowledge embedding extracted from an external knowledge base called ConceptNet. Specifically the proposed model employs a Graph Convolutional Neural Network(GCN) module to obtain commonsense knowledge embedding from the retrieved ConceptNet knowledge graph. By conducting a series of experiments with the VCR benchmark dataset, we show that the proposed KG_VCR model outperforms both the state of the art(SOTA) VQA model and the R2C VCR model.

• KIISE Transactions on Computing Practices
• /
• v.24 no.3
• /
• pp.138-144
• /
• 2018

A video is a vivid medium similar to human's visual-linguistic experiences, since it can inculcate a sequence of situations, actions or dialogues that can be told as a story. In this study, we propose story learning/regeneration frameworks from videos with successive event order supervision for contextual coherence. The supervision induces each episode to have a form of trajectory in the latent space, which constructs a composite representation of ordering and semantics. In this study, we incorporated the use of kids videos as a training data. Some of the advantages associated with the kids videos include omnibus style, simple/explicit storyline in short, chronological narrative order, and relatively limited number of characters and spatial environments. We build the encoder-decoder structure with successive event order embedding, and train bi-directional LSTMs as sequence models considering multi-step sequence prediction. Using a series of approximately 200 episodes of kids videos named 'Pororo the Little Penguin', we give empirical results for story regeneration tasks and SEOE. In addition, each episode shows a trajectory-like shape on the latent space of the model, which gives the geometric information for the sequence models.