TY - GEN
T1 - Automated Data Labeling for Object Detection via Iterative Instance Segmentation
AU - Kim, Jinyoon
AU - Kabir, Md Faisal
N1 - Publisher Copyright:
© 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Data labeling in computer vision, specifically in object detection tasks, remains a significant challenge in terms of efficiency and accuracy. This research article introduces an auto-labeling algorithm that combines active deep-learning techniques with the YOLOv8 model. The aim is to automate the data labeling process and enhance the performance of the object detection model. The proposed algorithm automatically labels a portion of unlabeled data based on uncertainty scores, integrating it into the training dataset. This approach reduces the need for manual annotation, which can be time-consuming. The effectiveness of the method is evaluated using two datasets: tomato and apple. The results demonstrate a substantial improvement in the Mean Average Precision score over multiple iterations, highlighting the enhanced performance of the overall model. Moreover, the experiments show that the proposed algorithm surpasses traditional manual annotation methods by generating a higher-performing model with significantly less annotation effort.
AB - Data labeling in computer vision, specifically in object detection tasks, remains a significant challenge in terms of efficiency and accuracy. This research article introduces an auto-labeling algorithm that combines active deep-learning techniques with the YOLOv8 model. The aim is to automate the data labeling process and enhance the performance of the object detection model. The proposed algorithm automatically labels a portion of unlabeled data based on uncertainty scores, integrating it into the training dataset. This approach reduces the need for manual annotation, which can be time-consuming. The effectiveness of the method is evaluated using two datasets: tomato and apple. The results demonstrate a substantial improvement in the Mean Average Precision score over multiple iterations, highlighting the enhanced performance of the overall model. Moreover, the experiments show that the proposed algorithm surpasses traditional manual annotation methods by generating a higher-performing model with significantly less annotation effort.
UR - http://www.scopus.com/inward/record.url?scp=85190101472&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85190101472&partnerID=8YFLogxK
U2 - 10.1109/ICMLA58977.2023.00124
DO - 10.1109/ICMLA58977.2023.00124
M3 - Conference contribution
AN - SCOPUS:85190101472
T3 - Proceedings - 22nd IEEE International Conference on Machine Learning and Applications, ICMLA 2023
SP - 845
EP - 850
BT - Proceedings - 22nd IEEE International Conference on Machine Learning and Applications, ICMLA 2023
A2 - Arif Wani, M.
A2 - Boicu, Mihai
A2 - Sayed-Mouchaweh, Moamar
A2 - Abreu, Pedro Henriques
A2 - Gama, Joao
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 22nd IEEE International Conference on Machine Learning and Applications, ICMLA 2023
Y2 - 15 December 2023 through 17 December 2023
ER -