http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
2차원/3차원 자유시점 비디오 재생을 위한 가상시점 합성시스템
민동보(Dongbo Min),손광훈(Kwanghoon Sohn) 대한전자공학회 2008 電子工學會論文誌-SP (Signal processing) Vol.45 No.4
3DTV를 위한 핵심 기술 중의 하나인 다시점 영상에서 변이를 추정하고 가상시점을 합성하는 새로운 방식을 제안한다. 다시점 영상에서 변이를 효율적이고 정확하게 추정하기 위해 준 N-시점 & N-깊이 구조를 제안한다. 이 구조는 이웃한 영상의 정보를 이용하여 변이 추정 시 발생하는 계산상의 중복을 줄인다. 제안 방식은 사용자에게 2D와 3D 자유시점을 제공하며, 사용자는 자유시점 비디오의 모드를 선택할 수 있다. 실험 결과는 제안 방식이 정확한 변이 지도를 제공하며, 합성된 영상이 사용자에게 자연스러운 자유시점 비디오를 제공한다는 것을 보여준다. In this paper, we propose a new approach for efficient multiview stereo matching and virtual view generation, which are key technologies for 3DTV. We propose semi N-view & N-depth framework to estimate disparity maps efficiently and correctly. This framework reduces the redundancy on disparity estimation by using the information of neighboring views. The proposed method provides a user 2D/3D freeview video, and the user can select 2D/3D modes of freeview video. Experimental results show that the proposed method yields the accurate disparity maps and the synthesized novel view is satisfactory enough to provide user seamless freeview videos.
마스크 이미지 모델링 가이드를 이용하여 개선된 스테레오 깊이 추정 기술 개발
안지혜,민동보 대한전자공학회 2024 대한전자공학회 학술대회 Vol.2024 No.6
In stereo matching tasks, CNN-based models have traditionally served as the predominant architectures, however, Transformer-based stereo models have also been adopted recently by leveraging effective pretraining methods to partially alleviate the inherent data-hungry issue in transformers. This paper focuses on addressing the labeled training data scarcity caused by the lack of locality inductive bias in the Transformer-based stereo models which is crucial for training with limited data when finetuning them in the downstream tasks such as stereo depth estimation. To mitigate this issue, we propose StereoIM, a novel stereo depth estimation framework that provides sufficient locality inductive biases during finetuning via Masked Image Modeling (MIM), which has been a prevalent approach for model pretraining. Reconstructing a masked image and subsequently predicting a disparity map from it, however, poses additional challenges in terms of stable model training. To overcome these challenges, we propose to employ an auxiliary network (teacher) that is updated via Exponential Moving Average (EMA) in addition to an original stereo model (student), and to use its predictions as supervisions for distilling the knowledge to the student. Our method achieves state-of-the-art on the KITTI 2015.
이헌상,손광훈,민동보 한국멀티미디어학회 2020 멀티미디어학회논문지 Vol.23 No.2
Recently, deep-learning based methods for low-light image enhancement accomplish great success through supervised learning. However, they still suffer from the lack of sufficient training data due to difficulty of obtaining a large amount of low-/normal-light image pairs in real environments. In this paper, we propose an unsupervised learning approach for single low-light image enhancement using the bright channel prior (BCP), which gives the constraint that the brightest pixel in a small patch is likely to be close to 1. With this prior, pseudo ground-truth is first generated to establish an unsupervised loss function. The proposed enhancement network is then trained using the proposed unsupervised loss function. To the best of our knowledge, this is the first attempt that performs a low-light image enhancement through unsupervised learning. In addition, we introduce a self-attention map for preserving image details and naturalness in the enhanced result. We validate the proposed method on various public datasets, demonstrating that our method achieves competitive performance over state-of-the-arts.
구재원(Jaywon Koo),민동보(Dongbo Min) 대한전자공학회 2021 대한전자공학회 학술대회 Vol.2021 No.6
While stereo matching based on deep networks has shown impressive results in daytime images, the performance is significantly degraded in nighttime images due to the lack of training data with ground truth and poor illumination condition. To overcome these issues, numerous methods have been proposed based on image-to-image translation. These approaches, however, often fail to predict accurate depth maps, when a domain gap between source (daytime) and target (nighttime) domains becomes large. In this paper, we propose a novel method for nighttime stereo matching to resolve such a performance degradation of the existing methods by a large domain gap. The large domain gap that often occurs between the day and night images is addressed using a two-step approach that consists of the image-to-image translation and domain adaptation. By utilizing additional pair of nighttime and daytime datasets which have smaller domain gap, our proposed model learns better image-to-image translation networks while jointly trained two domain adaptation networks explore to adapt to domains that have large domain gap. Extensive experiments on various datasets demonstrate that the proposed method outperforms state-of-the-arts approaches for nighttime stereo matching with a meaningful margin.