ETRI Journal最新文献

M. Yu, Y. Kwon, J. Lee, J. Park, J. Park, and T. Kim, PartitionTuner: an operator scheduler for deep-learning compilers supporting multiple heterogeneous processing units, ETRI Journal 45 (2023), 318–328. https://doi.org/10.4218/etrij.2021-0446

In the article entitled “PartitionTuner: An Operator Scheduler for Deep-Learning Compilers Supporting Multiple Heterogeneous Processing Units,” the authors would like to correct the caption of Figure 1, a typographical error in Section 2, and header of Table 2. The corrected text is provided below:

A typographical error appears on page 319. The correct version should read as:

That is, TVM's operator-scheduling technique enforces type-based backend mapping and sequential execution policy (TypeSeq).

A typographical error appears on page 320. The correct version should read as:

Figure 1B shows the final inference time of the DNN based on the execution time of each operator instead of the TVM's type-based backend mapping (OpSeq).

A typographical error appears in the caption of Figure 1 on page 320. The correct caption for Figure 1 should read as:

FIGURE 1 DNN inference time for different backend-mapping policies: (A) TypeSeq: static type-based backend mapping + sequential execution, (B) OpSeq: backend mapping based on the execution time of individual operator + sequential execution, and (C) PartitionTuner: backend mapping based on the execution time of individual/grouped operators + parallel execution

A typographical error appears in Table 2 on page 324. The unit 'Time (μs)' appearing in the table is incorrect. The correct should read as:

Time (ms)

The authors would like to apologize for the inconvenience caused.

T. An, J. Kang, D. Choi, and K.-W. Min, CRFNet: Context ReFinement Network used for semantic segmentation, ETRI Journal 45 (2023), 822–835. DOI 10.4218/etrij.2023-0017

In the article entitled “CRFNet: Context ReFinement Network used for semantic segmentation,” the authors would like to correct the author name of their article. “Taeghyun An” should be corrected to “Taeg-Hyun An.” The author's name has also been corrected in the online article.

The authors would like to apologize for the inconvenience caused.

RETRACTION: H. Nguyen. Van, N. Phan Minh, A. Ho Quoc Thien, and N. Nguyen Bao Thuc, “TransField: Improving Transformer Efficiency and Performance Through Conditional Random Fields,” ETRI Journal (Early View): https://doi.org/10.4218/etrij.2024-0491.

The above article, published online on 11 December 2024 in Wiley Online Library (wileyonlinelibrary.com), has been retracted by agreement between the authors; the journal Editor-in-Chief, Sungwon Yi, and John Wiley & Sons Australia, Ltd. The retraction has been agreed due to significant unattributed overlap between this article and a previously published article.[1]

REFERENCE

[1] H. Wu and K. Tu, “Probabilistic Transformer: A Probabilistic Dependency Model for Contextual Word Representation,” Findings of the Association for Computational Linguistics: ACL 2023 (2023): 7613–7636, https://doi.org/10.18653/v1/2023.findings-acl.482.

Website fingerprinting attacks that analyze clients' browsing preferences are an important information source for maintaining cybersecurity and improving big data utilization in terms of service quality. Many scholars have studied graph learning because of its powerful learning and inferential capabilities on unstructured network traffic. However, the extensive computing resource requirements of graph neural networks limit their application to large-scale and automated scenarios. In this paper, we propose a lightweight graph learning method for website fingerprinting attacks called GL-WF. We designed a four-level heterogeneous spatiotemporal graph using a multimodal representation of website browsing traffic. We then utilized a graph sampling algorithm to refine the structure of the graph. Finally, a well-designed graph isomorphism network was used to extract the graph topology for the traffic classifiers. The proposed GL-WF model, which was evaluated on over 3.6 million website flows, was found to provide high accuracy and efficiency.

This study investigated a noncoherent multiuser single-input multiple-output (SIMO) uplink transmission system based on large-scale array partitioning. In practical cases, limitations on the number of array elements and noncoherent detection make it difficult to resist nonlinear mutual interference and implement space division multiple access. In this study, based on energy difference detection (EDD), an array partitioning scheme and amplitude constellation were combined to improve the capacity of a massive SIMO system. EDD can suppress nonlinear mutual interference in energy-based SIMO, while the joint partitioning and amplitude constellation achieve both large-scale array diversity gain and spatial multi-access effects. Multi-partition detection can also be applied to cell-free SIMO scenarios. The complexity and overhead of the proposed scheme are tractable. Theoretical analysis demonstrates the feasibility of our scheme, and simulations reveal that the array partitioning scheme offers higher system capacity compared with non-partitioning schemes.

The development of smart cities depends on intelligent systems that integrate data from diverse environments. In this work, we present ELiOT, an end-to-end LiDAR odometry framework with transformer architecture designed to utilize real-world data, simulations, and digital twins. ELiOT leverages high-fidelity simulators and digital twin environments to enable sim-to-real applications, training on the real-world KITTI odometry dataset while benefiting from simulated data for improved generalization. Our self-attention-based flow embedding network eliminates the need for traditional 3D-2D projections by implicitly modeling motion from sequential LiDAR scans. The framework incorporates a 3D transformer encoder-decoder to extract rich geometric and semantic features. By integrating digital twin environments and simulated data into the training process, ELiOT bridges the gap between simulation and real-world applications, offering robust and scalable solutions for urban navigation challenges. This work underscores the potential of combining real-world and virtual data to advance LiDAR odometry and highlights its role for the future smart cities.

Advancements in autonomous vehicles and smart traffic systems require vision datasets capable of capturing complex interactions and dynamic behaviors in real-world urban environments. Although datasets such as COCO, Cityscapes, and ROAD have advanced object detection, segmentation, and action recognition, they often treat scene elements in isolation, thereby limiting their use for comprehensive understanding. This paper presents DOROS, a dataset with multilevel annotations across Agent, Location, and Behavior categories. DOROS is designed to support compositional reasoning under diverse traffic conditions. An annotation pipeline combining foundation models with structured human refinement ensures consistent, high-quality supervision. To support structured evaluation, we introduce the Combined mAP(mask) metric, which assesses instance segmentation under strict category-level label matching while mitigating the effects of class imbalance. Extensive experiments, including ablation studies and transformer-based baselines, validate DOROS as a resource for structured scene understanding in complex traffic scenarios. The dataset and code will be released upon publication.

This work tests the hypothesis that the primary bottleneck for visual quality in virtual try-on (VTON) systems is the precision of input segmentation masks, rather than generative capability. VTON technology empowers users to dress digital models in desired clothing items virtually. Conventional VTON models rely on segmentation models to isolate clothing regions and diffusion models to synthesize complete VTON images. This paper introduces high-speed and precise VTON (HSP-VTON) as a framework that uniquely combines refined two-stage semantic segmentation for enhanced accuracy with a latent consistency model to accelerate the diffusion-based image generation process. The synergistic integration of these components for VTON addresses critical challenges in both precision and speed. Experimental results on the ATR dataset demonstrate a 2.8% improvement in mean intersection over union compared with existing methods. Furthermore, HSP-VTON achieves superior performance on the VITON-HD dataset, outperforming state-of-the-art VTON models. The latent consistency model also reduces the number of inference steps, leading to substantial time savings without compromising image quality.