Temporal Action Proposal Generation with Background Constraint

Authors: Haosen Yang, Wenhao Wu, Lining Wang, Sheng Jin, Boyang Xia, Hongxun Yao, Hujie Huang3054-3062

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted on two popular benchmarks, i.e., Activity Net-1.3 and THUMOS14. The results demonstrate that our method outperforms state-of-the-art methods. Experiments Datasets and Evaluation Metrics
Researcher Affiliation Collaboration 1 Harbin Institute of Technology 2 Department of Computer Vision Technology (VIS), Baidu Inc. 3 University of Chinese Academy of Sciences
Pseudocode No The paper describes the model architecture and modules in text and diagrams, but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper refers to 'publicly available code' for existing methods (BMN, GTAD) but does not provide an explicit statement or link confirming that the code for the described methodology (BCNet) is publicly available.
Open Datasets Yes Activity Net-v1.3 is a large-scale video dataset for action recognition and temporal action detection tasks. It contains 10K training, 5k validation, and 5k testing videos with 200 action categories, and the ratio of training, validation and testing sets is 2:1:1. THUMOS14 contains 200 validation untrimmed videos and 213 test untrimmed videos, including 200 action categories.
Dataset Splits Yes It contains 10K training, 5k validation, and 5k testing videos with 200 action categories, and the ratio of training, validation and testing sets is 2:1:1.
Hardware Specification Yes (a single NVIDIA 2080Ti GPU).
Software Dependencies No The paper mentions using Adam optimizer but does not specify version numbers for other key software components, libraries, or programming languages like Python or PyTorch.
Experiment Setup Yes The number of layers in Boundary Prediction module is 12. For each anchor, we use sampling points N = 32. For postprocessing module, we set adjustment thresholds α1 = 0.9 and α2 = 0.8. We train our model from scratch using the Adam optimizer and the learning rate is set to 10 4 and decayed by a factor of 0.1 after every 10 epoch.