http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
황주비(Jubi Hwang),구인용(Inyong Koo),김창익(Changick Kim) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.11
Skin cancer often goes unnoticed by patients, making early diagnosis critical to prevent metastasis. Since dermatologists play a key role in determining diagnostic accuracy, there is a growing demand for a skin cancer diagnosis assistance system to support the process of discriminating and classifying malignant tumors. Recent advancements in deep learning have shown promise in medical image recognition.However, real-world clinical data presents a long-tailed distribution, which results in biased performance toward head classes, making it difficult to be introduced at the medical site. To address these challenges, we investigate the potential of employing a contrastive visual-language model in long-tailed skin lesion recognition. Through the utilization of both the visual features extracted from images and the semantic features from the skin lesion classes, our model can successfully mitigate the biasing issue arising from data imbalance. Additionally, we evaluate whether the conventional prompt is suitable for use in medical data and design prompts that reflect the expert knowledge extracted with ChatGPT.
다중 물체 추적을 위한 의사 깊이 기반 특징 벡터 갱신 알고리즘
심규진(Kyujin Shim),고강욱(Kangwook Ko),황주비(Jubi Hwang),김창익(Changick Kim) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.11
Multi-Object Tracking (MOT) is an essential task in computer vision, applied various fields like video surveillance and autonomous driving. Recently, tracking-by-detection-based trackers have shown promising performance and have become predominant approaches in MOT. Our study presents a novel pseudo-depth-based appearance feature vector update algorithm to improve those tracking-by-detection-based multi-object trackers. Experiments on the DanceTrack dataset reveal significant performance improvements compared to the previous feature updating strategies.
고강욱(Kangwook Ko),김희선(Hee-seon Kim),심규진(Kyujin Shim),황주비(Jubi Hwang),김창익(Changick Kim) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.11
This paper addresses the critical problem of Multiple Object Tracking (MOT) in computer vision. While traditional methods have primarily focused on pedestrian tracking, recent advances in the Transformer-based DETR (DEtection TRansformer) architecture, applied to object detection, have opened new possibilities for MOT. We introduce an approach that incorporates temporal characteristics into object tracking, utilizing moving averages in the Transformers encoder for image feature maps and decoder for tracking queries. Our method demonstrates improved performance, particularly in query mix, on the challenging DanceTrack dataset. By enhancing MOTRv2 with these temporal features, we pave the way for more effective and robust multiple object tracking solutions.