Weakly supervised video anomaly detection aims to detect anomalies and identify abnormal categories with only video-level labels. We propose CPL-VAD, a dual-branch framework with cross pseudo labeling. The binary anomaly detection branch focuses on snippet-level anomaly localization, while the category classification branch leverages vision-language alignment to recognize abnormal event categories. By exchanging pseudo labels, the two branches transfer complementary strengths, combining temporal precision with semantic discrimination. Experiments on XD-Violence and UCF-Crime demonstrate that CPL-VAD achieves state-of-the-art performance in both anomaly detection and abnormal category classification.
Figures & tables
Figure 1: Cross pseudo labeling framework.
Figure 2: The framework of CPL-VAD and Consistency-Aware Refinement module.
Figure 3: Qualitative results of CPL-VAD on XD-Violence.
Dataset@Metric
Method
XD@AP(%)
UCF@AUC(%)
MSLNET [ 4 ]
78.59
85.62
Zhang et al. [ 15 ]
78.74
86.22
OVVAD [ 12 ]
66.53
86.40
CLIP-TSA [ 2 ]
82.19
87.58
IFS-VAD [ 18 ]
83.14
86.57
Table 1: Coarse-grained comparisons on XD-Violence and UCF-Crime.
mAP@IoU
Method
0.1
0.2
0.3
0.4
0.5
Avg
RAD [ 7 ]
22.72
15.57
9.98
6.20
3.78
11.65
AVVD [ 11 ]
30.51
25.75
20.18
14.83
9.79
20.21
XD
VadCLIP [ 13 ]
37.03
30.84
23.38
17.90
14.31
24.70
CPL-VAD
45.89
40.28
33.49
27.59
20.40
33.53
RAD [ 7 ]
5.73
4.41
2.69
1.93
1.44
3.24
Table 2: Fine-grained comparisons on XD-Violence and UCF-Crime.
Pseudo label
Dataset
B → C
C → B
self
CAR
XD@AP, mAP(%)
UCF@AUC, mAP(%)
76.50, 29.97
86.60, 6.68
✓
✓
83.70, 31.47
87.34, 7.43
✓
✓
83.30, 32.43
87.04, 7.29
✓
✓
88.28, 27.75
87.61, 6.79
✓
✓
78.91, 31.86
86.42, 8.30
Table 3: Ablation studies of pseudo label structure and consistency-aware refinement module.
Figure 4: Comparison of qualitative results without and with cross pseudo labeling.