Learning Event Completeness for Weakly Supervised Video Anomaly Detection

Authors: Yu Wang, Shiwei Chen

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our LEC-VAD demonstrates remarkable advancements over the current state-of-the-art methods on two benchmark datasets XD-Violence and UCF-Crime. Extensive evaluations on the XD-Violence and UCF-Crime datasets have shown that our LEC-VAD achieves state-of-the-art performance.
Researcher Affiliation Collaboration 1School of Computer Science and Technology, Tongji University, Shanghai, China. 2Department of R&D Data, Microsoft Asia-Pacific Technology CO Ltd, Shanghai, China.. Correspondence to: Yu Wang <EMAIL>.
Pseudocode No The paper describes the methodology using textual explanations and mathematical equations, but it does not contain any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an explicit statement about the release of source code, nor does it provide any links to a code repository or mention code in supplementary materials.
Open Datasets Yes Datasets. UCF-Crime (Sultani et al., 2018) consists of 1900 untrimmed videos... XD-Violence (Wu et al., 2020) is a larger-scale benchmark comprising 4754 untrimmed videos...
Dataset Splits Yes UCF-Crime ... We adhere to a standard data split, where the training set and testing set comprise 1610 and 290 videos, respectively. XD-Violence ... 3954 videos and 800 videos are employed for training and testing respectively.
Hardware Specification No The paper mentions using 'pre-trained image encoders' and 'multiple vision encoders including I3D (Carreira & Zisserman, 2017), C3D (Tran et al., 2015), and the CLIP (Vi T-B/16)' to extract features, but does not specify the hardware (e.g., GPU, CPU models) used for training or running the experiments.
Software Dependencies No The paper mentions using a 'pre-trained text encoder of CLIP (Vi T-B/16)' and the 'Adam W optimizer,' but it does not provide specific version numbers for software libraries, programming languages, or frameworks used for implementation.
Experiment Setup Yes Implementation Details. The value of K is determined as K = max( T/16 , 1), and the momentum coefficient η is set to 0.99. We adopt the Adam W optimizer and train our LEC-VAD with a batch size of 64. The learning rate is set to 3e-5 and the model is trained for 10 epochs. We apply NMS with an Io U threshold of 0.5, and set the threshold rcls, and rano to 0.1 and 0.2. The hyper-parameters β, λ, γ, and m are explored in the experimental sections (Figure 5).