Towards Multi-Domain Learning for Generalizable Video Anomaly Detection

Authors: MyeongAh Cho, Taeoh Kim, Minho Shim, Dongyoon Wee, Sangyoun Lee

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experimental Results
Researcher Affiliation Collaboration Myeong Ah Cho Kyung Hee University maycho@khu.ac.kr Taeoh Kim NAVER Cloud taeoh.kim@navercorp.com Minho Shim NAVER Cloud minho.shim@navercorp.com Dongyoon Wee NAVER Cloud dongyoon.wee@navercorp.com Sangyoun Lee Yonsei University syleee@yonsei.ac.kr
Pseudocode No The paper includes 'Table A4: Illustration of each layer of proposed model' which details network architecture, but no structured pseudocode or algorithm blocks are provided.
Open Source Code No We will release the data and code when the paper is accepted.
Open Datasets Yes In this paper, we use six representative VAD datasets: UCF-Crimes (UCFC) [43], XD-Violences (XD) [51], Large-scale Anomaly Detection (LAD) [47], UBI-Fights (UBIF) [9], Traffic Anomaly Dataset (TAD) [19], and Shanghai-Tech Campus (ST) [24].
Dataset Splits Yes To achieve this, we sampled each dataset to align with the dataset with the smallest volume, ensuring that each abnormal category has a similar proportion and conducted 3-fold experiments.
Hardware Specification Yes All models and experiments are implemented and evaluated end-to-end using Py Torch [34] with a single NVIDIA V100 GPU.
Software Dependencies No All models and experiments are implemented and evaluated end-to-end using Py Torch [34] with a single NVIDIA V100 GPU.
Experiment Setup Yes The hyper-parameters are T = 32, λ = 10, τ = 0.3, and m = 0.3. Since the single-head baseline cannot assign AC labels using Eq. 6, pseudo-labels for the AC classifier are assigned based on the range of predicted abnormal scores.