TY - GEN
T1 - ML-VehicleDet
T2 - 15th International Conference on Digital Image Processing, ICDIP 2023
AU - Guan, Runwei
AU - Yao, Shanliang
AU - Zhu, Xiaohui
AU - Smith, Jeremy
AU - Man, Ka Lok
AU - Yue, Yutao
N1 - Publisher Copyright:
© 2023 ACM.
PY - 2023/5/19
Y1 - 2023/5/19
N2 - Vehicle detection based on deep learning has been developed rapidly and basically formed a certain pattern. Almost all works in vehicle detection are concentrated on single-label object detection. However, in the real world, a vehicle has multiple attributes from the perspective of a human being. When we observe a car, we tend to perceive its type, color, orientation, and other attributes. The categories of these labels are neither relevant nor hierarchical. For a neural network, this means that it needs to output multiple labels of a vehicle while regressing its bounding box. For traffic supervisors, a multi-label vehicle detection system could help them find out the targeted vehicle efficiently. Moreover, there has been no research on multi-label vehicle detection so far. Therefore, we design and develop a unified multi-label vehicle detection framework called ML-VehicleDet, which can detect the location (bounding box), type, color and orientation of the vehicle at the same time. In ML-VehicleDet, firstly, we design a hybrid one-stage object detection NN with ViT-encoder and CNN-decoder called Swin Only Look Once, which abbreviates SOLO. Such a SOLO is an anchor-free detector. Secondly, we design a practical loss function framework called MLC Loss, which includes two loss functions namely MLC-OM and MLC-OO for two different annotations of multi-label detection, specialized for alleviating the mutual inhibition problem in multi-label classification. Thirdly, we design a low-cost NMS algorithm called ML-NMS, specialized to merge bounding boxes with multiple labels for one vehicle. Furthermore, we reconstruct UA-DETRAC as a multi-label vehicle detection dataset (benchmark), called UA-DETRAC-ML. To the best of our knowledge, UA-DETRAC-ML is the first unified multi-label vehicle detection dataset. On UA-DETRAC-ML, ML-VehicleDet achieves 70.23% mAP, outperforming YOLOv5-M and YOLOX-M. To promote the development of the community, we release UA-DETRAC-ML at https://github.com/GuanRunwei/UA-DETRAC-ML.
AB - Vehicle detection based on deep learning has been developed rapidly and basically formed a certain pattern. Almost all works in vehicle detection are concentrated on single-label object detection. However, in the real world, a vehicle has multiple attributes from the perspective of a human being. When we observe a car, we tend to perceive its type, color, orientation, and other attributes. The categories of these labels are neither relevant nor hierarchical. For a neural network, this means that it needs to output multiple labels of a vehicle while regressing its bounding box. For traffic supervisors, a multi-label vehicle detection system could help them find out the targeted vehicle efficiently. Moreover, there has been no research on multi-label vehicle detection so far. Therefore, we design and develop a unified multi-label vehicle detection framework called ML-VehicleDet, which can detect the location (bounding box), type, color and orientation of the vehicle at the same time. In ML-VehicleDet, firstly, we design a hybrid one-stage object detection NN with ViT-encoder and CNN-decoder called Swin Only Look Once, which abbreviates SOLO. Such a SOLO is an anchor-free detector. Secondly, we design a practical loss function framework called MLC Loss, which includes two loss functions namely MLC-OM and MLC-OO for two different annotations of multi-label detection, specialized for alleviating the mutual inhibition problem in multi-label classification. Thirdly, we design a low-cost NMS algorithm called ML-NMS, specialized to merge bounding boxes with multiple labels for one vehicle. Furthermore, we reconstruct UA-DETRAC as a multi-label vehicle detection dataset (benchmark), called UA-DETRAC-ML. To the best of our knowledge, UA-DETRAC-ML is the first unified multi-label vehicle detection dataset. On UA-DETRAC-ML, ML-VehicleDet achieves 70.23% mAP, outperforming YOLOv5-M and YOLOX-M. To promote the development of the community, we release UA-DETRAC-ML at https://github.com/GuanRunwei/UA-DETRAC-ML.
KW - Multi-label object detection
KW - hybrid neural network
KW - multi-label classification
KW - non-maximum suppression
UR - http://www.scopus.com/inward/record.url?scp=85179894202&partnerID=8YFLogxK
U2 - 10.1145/3604078.3604108
DO - 10.1145/3604078.3604108
M3 - Conference Proceeding
AN - SCOPUS:85179894202
T3 - ACM International Conference Proceeding Series
BT - Proceedings of the 15th International Conference on Digital Image Processing, ICDIP 2023
PB - Association for Computing Machinery
Y2 - 19 May 2023 through 22 May 2023
ER -