JooYoung Jang - Academia.edu (original) (raw)
Uploads
Papers by JooYoung Jang
arXiv (Cornell University), Jan 14, 2020
The consistency loss has the potential for semi-supervised learning in segmentation and localizat... more The consistency loss has the potential for semi-supervised learning in segmentation and localization with various recent problems as well as classification. The conventional methods based on consistency loss, however, are inefficiently applied to semantic segmentation task regarded as pixel-wise classification; the consistency loss is calculated only in matched pixel pairs and averaged out. To this end, we propose a structured consistency loss for semi-supervised semantic segmentation to promote the consistency by considering the inter-pixel correlation. Specifically, the proposed method dramatically improves the computational complexity in the collaboration with cutmix [26]. It is shown via experiments that the Cityscapes benchmark results with validation data and test data are 81.9mIoU and 83.84mIoU respectively. This ranks the first place on the pixel-level semantic labeling task of Cityscapes benchmark suite. To the best of our knowledge, this is the first study to achieve the state-of-the-art performance with semi-supervised learning approach in semantic segmentation which provides the insight of the applicability of the semi-supervised technique to the various types of learning-based systems.
The Journal of Korea Navigation Institute, 2015
This paper introduces development of the MATLAB GUI based software for generating GPS RINEX obser... more This paper introduces development of the MATLAB GUI based software for generating GPS RINEX observation file. The purpose of this software is to generate GPS measurements of reference station or dynamic user, which are similar to the real GPS receiver data, accurately and efficiently. This software includes two data generation modes. One is Precision mode which generates GPS measurements as accurate as possible using post-processing data. The other is Real-time mode which generates GPS measurements using GPS error modeling technique. GPS error sources are calculated on the basis of each data generation mode, and L1/L2 pseudorange, L1/L2 carrier phase, and Doppler measurements are produced. These generated GPS measurements are recorded in the RINEX observation version 3.0 file. Using received GPS data at real reference station, we analyzed three items to verify software reliability; measurement bias, rate of change, and noise level. Consequently, RMS error of measurement bias is about 0.7 m, and this verification results demonstrate that our software can generate relatively exact GPS measurements.
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Many recent semi-supervised learning (SSL) studies build teacher-student architecture and train t... more Many recent semi-supervised learning (SSL) studies build teacher-student architecture and train the student network by the generated supervisory signal from the teacher. Data augmentation strategy plays a significant role in the SSL framework since it is hard to create a weak-strong augmented input pair without losing label information. Especially when extending SSL to semi-supervised object detection (SSOD), many strong augmentation methodologies related to image geometry and interpolation-regularization are hard to utilize since they possibly hurt the location information of the bounding box in the object detection task. To address this, we introduce a simple yet effective data augmentation method, Mix/UnMix (MUM), which unmixes feature tiles for the mixed image tiles for the SSOD framework. Our proposed method makes mixed input image tiles and reconstructs them in the feature space. Thus, MUM can enjoy the interpolation-regularization effect from non-interpolated pseudo-labels and successfully generate a meaningful weak-strong pair. Furthermore, MUM can be easily equipped on top of various SSOD methods. Extensive experiments on MS-COCO and PASCAL VOC datasets demonstrate the superiority of MUM by consistently improving the mAP performance over the baseline in all the tested SSOD benchmark protocols.
arXiv (Cornell University), Jan 14, 2020
The consistency loss has the potential for semi-supervised learning in segmentation and localizat... more The consistency loss has the potential for semi-supervised learning in segmentation and localization with various recent problems as well as classification. The conventional methods based on consistency loss, however, are inefficiently applied to semantic segmentation task regarded as pixel-wise classification; the consistency loss is calculated only in matched pixel pairs and averaged out. To this end, we propose a structured consistency loss for semi-supervised semantic segmentation to promote the consistency by considering the inter-pixel correlation. Specifically, the proposed method dramatically improves the computational complexity in the collaboration with cutmix [26]. It is shown via experiments that the Cityscapes benchmark results with validation data and test data are 81.9mIoU and 83.84mIoU respectively. This ranks the first place on the pixel-level semantic labeling task of Cityscapes benchmark suite. To the best of our knowledge, this is the first study to achieve the state-of-the-art performance with semi-supervised learning approach in semantic segmentation which provides the insight of the applicability of the semi-supervised technique to the various types of learning-based systems.
The Journal of Korea Navigation Institute, 2015
This paper introduces development of the MATLAB GUI based software for generating GPS RINEX obser... more This paper introduces development of the MATLAB GUI based software for generating GPS RINEX observation file. The purpose of this software is to generate GPS measurements of reference station or dynamic user, which are similar to the real GPS receiver data, accurately and efficiently. This software includes two data generation modes. One is Precision mode which generates GPS measurements as accurate as possible using post-processing data. The other is Real-time mode which generates GPS measurements using GPS error modeling technique. GPS error sources are calculated on the basis of each data generation mode, and L1/L2 pseudorange, L1/L2 carrier phase, and Doppler measurements are produced. These generated GPS measurements are recorded in the RINEX observation version 3.0 file. Using received GPS data at real reference station, we analyzed three items to verify software reliability; measurement bias, rate of change, and noise level. Consequently, RMS error of measurement bias is about 0.7 m, and this verification results demonstrate that our software can generate relatively exact GPS measurements.
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Many recent semi-supervised learning (SSL) studies build teacher-student architecture and train t... more Many recent semi-supervised learning (SSL) studies build teacher-student architecture and train the student network by the generated supervisory signal from the teacher. Data augmentation strategy plays a significant role in the SSL framework since it is hard to create a weak-strong augmented input pair without losing label information. Especially when extending SSL to semi-supervised object detection (SSOD), many strong augmentation methodologies related to image geometry and interpolation-regularization are hard to utilize since they possibly hurt the location information of the bounding box in the object detection task. To address this, we introduce a simple yet effective data augmentation method, Mix/UnMix (MUM), which unmixes feature tiles for the mixed image tiles for the SSOD framework. Our proposed method makes mixed input image tiles and reconstructs them in the feature space. Thus, MUM can enjoy the interpolation-regularization effect from non-interpolated pseudo-labels and successfully generate a meaningful weak-strong pair. Furthermore, MUM can be easily equipped on top of various SSOD methods. Extensive experiments on MS-COCO and PASCAL VOC datasets demonstrate the superiority of MUM by consistently improving the mAP performance over the baseline in all the tested SSOD benchmark protocols.