Beyond Euclidean: Dual-Space Representation Learning for Weakly Supervised Video Violence Detection

Authors: Jiaxu Leng, Zhanjie Wu, Mingpi Tan, Yiran Liu, Ji Gan, Haosheng Chen, Xinbo Gao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments demonstrate the effectiveness of our proposed DSRL.
Researcher Affiliation Academia Jiaxu Leng1,2, Zhanjie Wu1,2, Mingpi Tan1,2, Yiran Liu1, Ji Gan1,2, Haosheng Chen1,2, Xinbo Gao 1,2 1 Chongqing University of Posts and Telecommunications, Chongqing, China 2 Chongqing Institute for Brain and Intelligence, Guangyang Bay Laboratory, Chongqing, China gaoxb@cqupt.edu.cn
Pseudocode No The paper describes the methodology through text and diagrams but does not include explicit pseudocode or algorithm blocks.
Open Source Code No We will open-source the code in the future.
Open Datasets Yes Datasets. Under the multimodal input setting, we follow [34, 37, 25] to conduct experiments on XD-Violence, which is the only and extremely challenging VVD dataset with multimodal information. Under the unimodal input setting, both the XD-Violence[34] and UCF-Crime[30] datasets are used to evaluate our method.
Dataset Splits No The paper mentions training and test sets (e.g., '1,610 training videos' and '290 test videos' for UCF-Crime) but does not explicitly describe a separate validation set split or how it was used.
Hardware Specification Yes We use an Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz, a NVIDIA RTX A6000 GPU to conduct experiments.
Software Dependencies Yes We use CUDA 12.2, Python 3.9.16, and Pytorch 1.12.1.
Experiment Setup Yes Our proposed method is trained for 30 epochs in total, and the batch size is 256. The initial learning rate is 0.001, which is dynamically adjusted by a cosine annealing scheduler [13]. We use Adam [14] as the optimizer without weight decay. For hyper-parameters, we set β as 0.8, γ as 1.2, α as 0.3, and dropout rate as 0.6.