Seung-Hwan Bae
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Seung-Hwan Bae has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
Seong-Ho Lee
3 shared publications- Dae-Hyeon Park
Vision and Learning Laboratory, Inha University, Incheon, Republic of Korea
2 shared publications - Deok-Woong Kim2 shared publications
- Seong-Ho Lee
Vision and Learning Laboratory, Inha University, Incheon, Republic of Korea
2 shared publications Yong-Sang Yoo
Department of Computer Science and Engineering, Incheon National University, 119 Academy-ro, Yeonsu-gu, Incheon, 22012, South Korea
2 shared publications- Chang Mook Kang
Department of Electrical Engineering, Hanyang University, Seoul 04763, Republic of Korea
1 shared publication
Latest Publications
- Open Access
- 16 April 2026
- Published by Elsevier in ICT Express
- Vol. 12 (3) , 707-713
Improving data-free quantization with confidence-guided data synthesis
- 16 April 2026
- Published by Elsevier in ICT Express
- Vol. 12 (3) , 707-713
Open AccessImproving data-free quantization with confidence-guided data synthesis
- Open Access
- 9 February 2026
- Published by Elsevier in ICT Express
- Vol. 12 (2) , 523-529
Gated side adapters with memory-efficient fine tuning for RGB-T tracking
- 9 February 2026
- Published by Elsevier in ICT Express
- Vol. 12 (2) , 523-529
Open AccessGated side adapters with memory-efficient fine tuning for RGB-T tracking
- Open Access
- 20 August 2025
- Published by Elsevier in ICT Express
- Vol. 11 (5) , 974-980
Neural-NGBoost: Natural gradient boosting with neural network base learners
- 20 August 2025
- Published by Elsevier in ICT Express
- Vol. 11 (5) , 974-980
Open AccessNeural-NGBoost: Natural gradient boosting with neural network base learners
- Open Access
- 19 May 2025
- Published by Elsevier in ICT Express
- Vol. 11 (4) , 696-702
Region-aware knowledge distillation between monocular camera-based 3D object detectors
- 19 May 2025
- Published by Elsevier in ICT Express
- Vol. 11 (4) , 696-702
Open AccessRegion-aware knowledge distillation between monocular camera-based 3D object detectors
- Open Access
- 11 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 53533-53548
Bridging the Knowledge Gap via Transformer-Based Multi-Layer Correlation Learning
We tackle a multi-layer knowledge distillation problem between deep models with heterogeneous architectures. The main challenges of that are the mismatches of the feature maps in terms of the resolution or semantic levels. To resolve this, we propose a novel transformer-based multi-layer correlation knowledge distillation (TMC-KD) method in order to bridge the knowledge gap between a pair of networks. Our method aims to narrow the relational knowledge gaps between teacher and student models by minimizing the local and global feature correlations. Based on extensive comparisons with the recent KD methods on classification and detection tasks, we prove the effectiveness and usefulness of our TMC-KD method.
- 11 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 53533-53548
Open AccessBridging the Knowledge Gap via Transformer-Based Multi-Layer Correlation Learning
We tackle a multi-layer knowledge distillation problem between deep models with heterogeneous architectures. The main challenges of that are the mismatches of the feature maps in terms of the resolution or semantic levels. To resolve this, we propose a novel transformer-based multi-layer correlation knowledge distillation (TMC-KD) method in order to bridge the knowledge gap between a pair of networks. Our method aims to narrow the relational knowledge gaps between teacher and student models by minimizing the local and global feature correlations. Based on extensive comparisons with the recent KD methods on classification and detection tasks, we prove the effectiveness and usefulness of our TMC-KD method.
- 28 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 32 (10577149) , 4378-4392
Decode-MOT: How Can We Hurdle Frames to Go Beyond Tracking-by-Detection?
The speed of tracking-by-detection (TBD) greatly depends on the number of running a detector because the detection is the most expensive operation in TBD. In many practical cases, multi-object tracking (MOT) can be, however, achieved based tracking-by-motion (TBM) only. This is a possible solution without much loss of MOT accuracy when the variations of object cardinality and motions are not much within consecutive frames. Therefore, the MOT problem can be transformed to find the best TBD and TBM mechanism. To achieve it, we propose a novel decision coordinator for MOT (Decode-MOT) which can determine the best TBD/TBM mechanism according to scene and tracking contexts. In specific, our Decode-MOT learns tracking and scene contextual similarities between frames. Because the contextual similarities can vary significantly according to the used trackers and tracking scenes, we learn the Decode-MOT via self-supervision. The evaluation results on MOT challenge datasets prove that our method can boost the tracking speed greatly while keeping the state-of-the-art MOT accuracy. Our code will be available at https://github.com/reussite-cv/Decode-MOT.
- 28 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 32 (10577149) , 4378-4392
Decode-MOT: How Can We Hurdle Frames to Go Beyond Tracking-by-Detection?
The speed of tracking-by-detection (TBD) greatly depends on the number of running a detector because the detection is the most expensive operation in TBD. In many practical cases, multi-object tracking (MOT) can be, however, achieved based tracking-by-motion (TBM) only. This is a possible solution without much loss of MOT accuracy when the variations of object cardinality and motions are not much within consecutive frames. Therefore, the MOT problem can be transformed to find the best TBD and TBM mechanism. To achieve it, we propose a novel decision coordinator for MOT (Decode-MOT) which can determine the best TBD/TBM mechanism according to scene and tracking contexts. In specific, our Decode-MOT learns tracking and scene contextual similarities between frames. Because the contextual similarities can vary significantly according to the used trackers and tracking scenes, we learn the Decode-MOT via self-supervision. The evaluation results on MOT challenge datasets prove that our method can boost the tracking speed greatly while keeping the state-of-the-art MOT accuracy. Our code will be available at https://github.com/reussite-cv/Decode-MOT.
- 20 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 45 (9) , 10817-10834
Deformable Part Region Learning and Feature Aggregation Tree Representation for Object Detection
Region-based object detection infers object regions for one or more categories in an image. Due to the recent advances in deep learning and region proposal methods, object detectors based on convolutional neural networks (CNNs) have been flourishing and provided promising detection results. However, the accuracy of the convolutional object detectors can be degraded often due to the low feature discriminability caused by geometric variation or transformation of an object. In this article, we propose a deformable part region (DPR) learning in order to allow decomposed part regions to be deformable according to the geometric transformation of an object. Because the ground truth of the part models is not available in many cases, we design part model losses for the detection and segmentation, and learn the geometric parameters by minimizing an integral loss including those part losses. As a result, we can train our DPR network without extra supervision, and make multi-part models deformable according to object geometric variation. Moreover, we propose a novel feature aggregation tree (FAT) so as to learn more discriminative region of interest (RoI) features via bottom-up tree construction. The FAT can learn the stronger semantic features by aggregating part RoI features along the bottom-up pathways of the tree. We also present a spatial and channel attention mechanism for the aggregation between different node features. Based on the proposed DPR and FAT networks, we design a new cascade architecture that can refine detection tasks iteratively. Without bells and whistles, we achieve impressive detection and segmentation results on MSCOCO and PASCAL VOC datasets. Our Cascade D-PRD achieves the 57.9 box AP with the Swin-L backbone. We also provide an extensive ablation study to prove the effectiveness and usefulness of the proposed methods for large-scale object detection.
- 20 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 45 (9) , 10817-10834
Deformable Part Region Learning and Feature Aggregation Tree Representation for Object Detection
Region-based object detection infers object regions for one or more categories in an image. Due to the recent advances in deep learning and region proposal methods, object detectors based on convolutional neural networks (CNNs) have been flourishing and provided promising detection results. However, the accuracy of the convolutional object detectors can be degraded often due to the low feature discriminability caused by geometric variation or transformation of an object. In this article, we propose a deformable part region (DPR) learning in order to allow decomposed part regions to be deformable according to the geometric transformation of an object. Because the ground truth of the part models is not available in many cases, we design part model losses for the detection and segmentation, and learn the geometric parameters by minimizing an integral loss including those part losses. As a result, we can train our DPR network without extra supervision, and make multi-part models deformable according to object geometric variation. Moreover, we propose a novel feature aggregation tree (FAT) so as to learn more discriminative region of interest (RoI) features via bottom-up tree construction. The FAT can learn the stronger semantic features by aggregating part RoI features along the bottom-up pathways of the tree. We also present a spatial and channel attention mechanism for the aggregation between different node features. Based on the proposed DPR and FAT networks, we design a new cascade architecture that can refine detection tasks iteratively. Without bells and whistles, we achieve impressive detection and segmentation results on MSCOCO and PASCAL VOC datasets. Our Cascade D-PRD achieves the 57.9 box AP with the Swin-L backbone. We also provide an extensive ablation study to prove the effectiveness and usefulness of the proposed methods for large-scale object detection.
- 27 January 2023
- Published by Elsevier in Pattern Recognition
- Vol. 138
AFI-GAN: Improving feature interpolation of feature pyramid networks via adversarial training for object detection
- 27 January 2023
- Published by Elsevier in Pattern Recognition
- Vol. 138
AFI-GAN: Improving feature interpolation of feature pyramid networks via adversarial training for object detection
- Open Access
- 1 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 318-330
RT-MOT: Confidence-Aware Real-Time Scheduling Framework for Multi-Object Tracking Tasks
Different from existing MOT (Multi-Object Tracking) techniques that usually aim at improving tracking accuracy and average FPS, real-time systems such as autonomous vehicles necessitate new requirements of MOT under limited computing resources: (R1) guarantee of timely execution and (R2) high tracking accuracy. In this paper, we propose RT-MOT, a novel system design for multiple MOT tasks, which addresses R1 and R2. Focusing on multiple choices of a workload pair of detection and association, which are two main components of the tracking-by-detection approach for MOT, we tailor a measure of object confidence for RT-MOT and develop how to estimate the measure for the next frame of each MOT task. By utilizing the estimation, we make it possible to predict tracking accuracy variation according to different workload pairs to be applied to the next frame of an MOT task. Next, we develop a novel confidence-aware real-time scheduling framework, which offers an offline timing guarantee for a set of MOT tasks based on non-preemptive fixed-priority scheduling with the smallest workload pair. At run-time, the framework checks the feasibility of a priority-inversion associated with a larger workload pair, which does not compromise the timing guarantee of every task, and then chooses a feasible scenario that yields the largest tracking accuracy improvement based on the proposed prediction. Our experiment results demonstrate that RT-MOT significantly improves overall tracking accuracy by up to 1.5 ×, compared to existing popular tracking-by-detection approaches, while guaranteeing timely execution of all MOT tasks.
- 1 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 318-330
Open AccessRT-MOT: Confidence-Aware Real-Time Scheduling Framework for Multi-Object Tracking Tasks
Different from existing MOT (Multi-Object Tracking) techniques that usually aim at improving tracking accuracy and average FPS, real-time systems such as autonomous vehicles necessitate new requirements of MOT under limited computing resources: (R1) guarantee of timely execution and (R2) high tracking accuracy. In this paper, we propose RT-MOT, a novel system design for multiple MOT tasks, which addresses R1 and R2. Focusing on multiple choices of a workload pair of detection and association, which are two main components of the tracking-by-detection approach for MOT, we tailor a measure of object confidence for RT-MOT and develop how to estimate the measure for the next frame of each MOT task. By utilizing the estimation, we make it possible to predict tracking accuracy variation according to different workload pairs to be applied to the next frame of an MOT task. Next, we develop a novel confidence-aware real-time scheduling framework, which offers an offline timing guarantee for a set of MOT tasks based on non-preemptive fixed-priority scheduling with the smallest workload pair. At run-time, the framework checks the feasibility of a priority-inversion associated with a larger workload pair, which does not compromise the timing guarantee of every task, and then chooses a feasible scenario that yields the largest tracking accuracy improvement based on the proposed prediction. Our experiment results demonstrate that RT-MOT significantly improves overall tracking accuracy by up to 1.5 ×, compared to existing popular tracking-by-detection approaches, while guaranteeing timely execution of all MOT tasks.
- Open Access
Effective Multi-Object Tracking via Global Object Models and Object Constraint Learning
Effective multi-object tracking is still challenging due to the trade-off between tracking accuracy and speed. Because the recent multi-object tracking (MOT) methods leverage object appearance and motion models so as to associate detections between consecutive frames, the key for effective multi-object tracking is to reduce the computational complexity of learning both models. To this end, this work proposes global appearance and motion models to discriminate multiple objects instead of learning local object-specific models. In concrete detail, it learns a global appearance model using contrastive learning between object appearances. In addition, we learn a global relation motion model using relative motion learning between objects. Moreover, this paper proposes object constraint learning for improving tracking efficiency. This study considers the discriminability of the models as a constraint, and learns both models when inconsistency with the constraint occurs. Therefore, object constraint learning differs from the conventional online learning for multi-object tracking which updates learnable parameters per frame. This work incorporates global models and object constraint learning into the confidence-based association method, and compare our tracker with the state-of-the-art methods on public available MOT Challenge datasets. As a result, we achieve 64.5% MOTA (multi-object tracking accuracy) and 6.54 Hz tracking speed on the MOT16 test dataset. The comparison results show that our methods can contribute to improve tracking accuracy and tracking speed together.
Open AccessEffective Multi-Object Tracking via Global Object Models and Object Constraint Learning
Effective multi-object tracking is still challenging due to the trade-off between tracking accuracy and speed. Because the recent multi-object tracking (MOT) methods leverage object appearance and motion models so as to associate detections between consecutive frames, the key for effective multi-object tracking is to reduce the computational complexity of learning both models. To this end, this work proposes global appearance and motion models to discriminate multiple objects instead of learning local object-specific models. In concrete detail, it learns a global appearance model using contrastive learning between object appearances. In addition, we learn a global relation motion model using relative motion learning between objects. Moreover, this paper proposes object constraint learning for improving tracking efficiency. This study considers the discriminability of the models as a constraint, and learns both models when inconsistency with the constraint occurs. Therefore, object constraint learning differs from the conventional online learning for multi-object tracking which updates learnable parameters per frame. This work incorporates global models and object constraint learning into the confidence-based association method, and compare our tracker with the state-of-the-art methods on public available MOT Challenge datasets. As a result, we achieve 64.5% MOTA (multi-object tracking accuracy) and 6.54 Hz tracking speed on the MOT16 test dataset. The comparison results show that our methods can contribute to improve tracking accuracy and tracking speed together.