TY - GEN
T1 - WSAD-Net
T2 - 12th International Conference on Image and Graphics, ICIG 2023
AU - Wu, Peng
AU - Zhang, Yanning
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.
PY - 2023
Y1 - 2023
N2 - Weakly supervised anomaly detection (WSAD) is a newfangled and challenging task, the goal of which is to detect anomalous activities in untrimmed surveillance videos with no requirement of temporal localization annotations. A few methods have been proposed to detect anomalies under the weakly supervised setting. In order to combat the issue even further, we present a weakly supervised anomaly detector network (WSAD-Net), which is composed of a pre-trained feature extractor and an anomaly-specific subnetwork. To learn the anomaly-specific parameters of WSAD-Net, we design a Classification Loss based on the multiple instance learning (MIL) and two novel losses, namely, Compactness Loss and Magnetism Loss, which play an important role in evaluating the correlation of features. During the test phase, we introduce the Anomaly-specific Temporal Class Activation Sequence (Ano-TCAS) to generate the anomaly score. We evaluate WSAD-Net on two benchmarks, i.e., the UCF-Crime and Live-Videos datasets, and experiments on these two benchmarks show that WSAD-Net outperforms or competes with current state-of-the-art methods.
AB - Weakly supervised anomaly detection (WSAD) is a newfangled and challenging task, the goal of which is to detect anomalous activities in untrimmed surveillance videos with no requirement of temporal localization annotations. A few methods have been proposed to detect anomalies under the weakly supervised setting. In order to combat the issue even further, we present a weakly supervised anomaly detector network (WSAD-Net), which is composed of a pre-trained feature extractor and an anomaly-specific subnetwork. To learn the anomaly-specific parameters of WSAD-Net, we design a Classification Loss based on the multiple instance learning (MIL) and two novel losses, namely, Compactness Loss and Magnetism Loss, which play an important role in evaluating the correlation of features. During the test phase, we introduce the Anomaly-specific Temporal Class Activation Sequence (Ano-TCAS) to generate the anomaly score. We evaluate WSAD-Net on two benchmarks, i.e., the UCF-Crime and Live-Videos datasets, and experiments on these two benchmarks show that WSAD-Net outperforms or competes with current state-of-the-art methods.
KW - Temporal Class Activation Sequence
KW - Untrimmed Surveillance Video
KW - Video Anomaly Detection
KW - Weak Supervision
UR - http://www.scopus.com/inward/record.url?scp=85177419425&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-46317-4_22
DO - 10.1007/978-3-031-46317-4_22
M3 - 会议稿件
AN - SCOPUS:85177419425
SN - 9783031463167
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 271
EP - 282
BT - Image and Graphics - 12th International Conference, ICIG 2023, Proceedings
A2 - Lu, Huchuan
A2 - Liu, Risheng
A2 - Ouyang, Wanli
A2 - Huang, Hui
A2 - Lu, Jiwen
A2 - Dong, Jing
A2 - Xu, Min
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 22 September 2023 through 24 September 2023
ER -
