YOLO 성능 향상을 위한 데이터 증강기법
Data Augmentation for the enhancement of YOLO Performance
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
컴퓨터 비전은 CNN, 트랜스포머 등과 같은 모델의 발전으로 여러 분야에서 좋은 성과를 이루었다. 하지만, 모델을 학습하기 위해서는 다양하고 많은 데이터가 필요하다. 이러한 학습데이터를 얻기 위해서는 많은 시간과 노력이 필요로 한다. 이러한 높은 비용으로 인해 데이터 부족이나 데이터 불균형이 발생하게 된다. 데이터 증강기법은 이러한 문제를 해결하기 위한 좋은 방법이다. 본 논문에서는 객체 인식 모델을 위한 데이터 증강기법 중에서 (복사-붙여넣기) Copy-Paste를 활용한 데이터 증강기법을 연구한다. 이전 연구에서는 인스턴스 영상 분할 객체를 붙이거나 시각적인 맥락을 바탕으로 객체를 붙인다. 하지만 인스턴스 영상 분할 객체를 사용하지 않고 단순한 방법인 바운딩 박스(Bounding Box)를 그대로 기존의 객체 위치에 같은 크기로 붙이거나 무작위로 붙이는 것도 모델의 성능이 향상된다는 것을 발견했다. 또한, 객체에서 SAM(Segment Anything Model) 모델을 활용하여 객체의 인스턴스를 추출하여 붙이는 방법을 제안한다. 그리고 붙이는 객체에 데이터 증강기법을 적용하여 데이터를 증강하는 방법을 추가실험으로 보여준다. 또한, 기존의 객체가 붙여지는 객체에 의해 가려지는 것을 막기 위해 객체를 붙이고 기존 이미지에 있는 객체를 덮어쓴 방법도 적용하였다. 본 논문에서 객체 인식 모델 Yolo v5를 Pascal VOC12 데이터셋으로만 학습한 결과보다 제안한 데이터 증강기법을 활용해서 학습한 결과가 더 높은 성능을 보여주는 것을 확인하였다.
Abstract
Computer vision has shown excellent performance in various fields, thanks to the advancements in models like CNN and Transformers. However, training these models requires diverse and abundant data, which demands a significant amount of time and effort. The high cost associated with acquiring such training data often leads to issues like data scarcity and data imbalance. Data augmentation techniques provide effective solutions to address these challenges. In this paper, we focus on researching data augmentation techniques for object recognition models, specifically leveraging the Copy-Paste (Augmentation) technique. The previous researches involved attaching objects based on instance segmentation or visual context. However, we have discovered that using a straightforward approach, such as attaching bounding boxes of the same size to the existing object locations or randomly attaching objects, enhances the model's performance significantly. Furthermore, we propose a method of using the SAM (Segment Anything Model) to extract object instances from images and attaching them. We demonstrate additional experiments applying data augmentation techniques to the attached objects. To prevent existing objects in the image from occluded by the attached objects, we present a method of overlaying them into the image with attached objects. In this paper, we train the object recognition model using YOLO (You Only Look Once) v5 on the Pascal VOC12 dataset, and show better performance when utilizing the proposed data augmentation techniques.
참고문헌 (44)
- D. N. Perkins, "Transfer of Learning" (1992), Pergamon
- S. J. Pan, "A survey on transfer learning", IEEE Trans. Knowl. Data Eng, 22(10), pp.1345~1359 (2010)
- Z. Wang, "Characterizing and avoiding negative transfer", Proc. IEEE Conference on Computer Vision and Pattern Recognition, pp.1129311302~ (2019)
- A. Dosovitskiy, "Discriminative Unsupervised Feature Learning with Exemplar Convolutional Neural Networks", NeurIPS (2014)
- X. Gao, "GraphTER : Unsupervised Learning of Graph Transformation Equivariant Representations via Auto-Encoding Node-wise Transformations", CVPR (2020)
- A. Kolesnikov, "Revisiting Self-Supervised Visual Representation Learning", CVPR (2019)
- Li, Q., "Age progression and regression with spatial attention modules", Proceedings of the AAAI Conference on Artificial Intelligence, pp.11378~11385 (2020)
- Zhu, J. Y., "Unpaired image-to-image translation using cycle-consistent adversarial networks", Proceedings of the IEEE international conference on computer vision, pp.2223~2232 (2017)
- Terrance DeVries, "DATASET AUGMENTATION IN FEATURE SPACE", ICLR (2017)
- Chia-Wen Kuo, "FeatMatch : Feature-Based Augmentation for Semi-Supervised Learning", ECCV (2020)
- Daniel Ho, "Population based augmentation:Efficient learning of augmentation policy schedules", Proc. Int. Conf. Mach. Learn, pp.2731~2741 (2019), PMLR
- Sungbin Lim, "Fast autoaugment", Proc. Adv. Neural Inf. Process. Syst, 32 (2019)
- Ekin D Cubuk, "Randaugment: Practical automated data augmentation with a reduced search space", Proc. IEEE Conf. Comput. Vis. Pattern Recog. Workshops, pp.702~703 (2020)
- Sangdoo Yun, "Cutmix : Regularization strategy to train strong classifiers with localizable features", Proceedings of the IEEE, 2019, 6023~6032
- Hongyi Zhang, "mixup: Beyond empirical risk minimization", International Conference on Learning Representations, pp.2018~ (2018)
- Nikita Dvornik, "Modeling visual context is key to augmenting object detection datasets", ECCV (2018)
- Hao-Shu Fang, "Instaboost: Boosting instance segmentation via probability map guided copypasting", ICCV (2019)
- Debidatta Dwibedi, "Cut, paste and learn: Surprisingly easy synthesis for instance detection" (2017), ICCV
- Golnaz Ghiasi, "Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation", CVPR (2021)
- Mark Everingham, "The pascal visual object classes (voc) challenge", IJCV (2010)
- Seulki Park, "The majority can help the minority: Contextrich minority oversampling for long-tailed classification", CVPR, pp.6887~6896 (2022), IEEE[DOI]
- Vikas Verma, "Manifold mixup : Better representations by interpolating hidden states", International Conference on Machine Learning, pp.6438~6447 (2019), PMLR
- Mojtaba Faramarzi, "Patchup: A feature-space block-level regularization technique for convolutional neural networks", AAAI, 36, pp.589~597 (2022)
- R. Girshick, "Rich feature hierarchies for accurate object detection and semantic segmentation", CVPR (2014)
- Ross Girshick, "Fast r-cnn", ICCV (2015)
- ShaoqingRen, "Faster r-cnn: Towards real-time object detection with region proposal networks", NeurIPS (2015)
- KaimingHe, "Mask r-cnn", ICCV (2017)
- Tsung-Yi Lin, "Focal loss for dense object detection", ICCV (2017)
- Wei Liu, "Ssd: Single shot multiboxdetector", ECCV (2016)
- Joseph Redmon, "You only look once: Unified, real-time object detection", CVPR (2016)
- Hei Law, "Cornernet : Detecting objects as paired keypoints", ECCV (2018)
- Zhi Tian, "Fcos: Fully convolutional one-stage object detection", ICCV (2019)
- Nicolas Carion, "End-toendobject detection with transformers", ECCV (2020)
- ZhigangDai, "Up-detr: Unsupervised pre-training for object detection with transformers", CVPR (2021)
- Feng Li, "Dn-detr:Accelerate detrtraining by introducing query denoising", CVPR (2022)
- XizhouZhu, "Deformable detr: Deformable transformers for end-to-end object detection", ICLR (2021)
- M. Everingham, , The PASCAL Visual Object Classes Challenge 2007(VOC2007) (2007)
- J. Long, "Fully convolutional networks for semantic segmentation", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.3431~3440 (2015)
- S. M. Pizer, "Contrastlimited adaptive histogram equalization : speed and effectiveness", Proceedings of the First Conference on Visualization in Biomedical Computing, pp.337~345 (1990)