Spatio-Temporal Deformable Convolution for Compressed Video Quality Enhancement

Authors: Jianing Deng, Li Wang, Shiliang Pu, Cheng Zhuo10696-10703

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our method achieves the state-of-the-art performance of compressed video quality enhancement in terms of both accuracy and efficiency.
Researcher Affiliation Collaboration Jianing Deng,1 Li Wang,2 Shiliang Pu,2 Cheng Zhuo1 1College of Information Science and Electronic Engineering, Zhejiang University 2Hikvision Research Institute {dengjn, czhuo}@zju.edu.cn, {wangli7, pushiliang}@hikvision.com
Pseudocode No The paper describes the proposed method using mathematical formulations and architectural diagrams but does not include explicit pseudocode or algorithm blocks.
Open Source Code No The paper mentions that the method is implemented based on PyTorch and references MMDetection toolbox, but it does not provide a link or an explicit statement about the release of its own source code.
Open Datasets Yes Following MFQE 2.0 (Guan et al. 2019), we collect a total of 130 uncompressed videos with various resolutions and contents from two databases, i.e., Xiph (Xiph.org) and VQEG (VQEG), where 106 of them are selected for training and the rest are for validation.
Dataset Splits Yes Following MFQE 2.0 (Guan et al. 2019), we collect a total of 130 uncompressed videos with various resolutions and contents from two databases, i.e., Xiph (Xiph.org) and VQEG (VQEG), where 106 of them are selected for training and the rest are for validation. For testing, we adopt the dataset from Joint Collaborative Team on Video Coding (Ohm et al. 2012) with 18 uncompressed videos.
Hardware Specification Yes Results of speed are measured on Nvidia Ge Force GTX 1080 Ti GPU.
Software Dependencies No The proposed method is implemented based on Py Torch framework with reference to MMDetection toolbox (Chen et al. 2019) for deformable convolution. ... We train all models using Adam optimizer (Kingma and Ba 2014). While software names are mentioned, specific version numbers for PyTorch and MMDetection are not provided.
Experiment Setup Yes For training, we randomly crop 64 64 clips from raw and the corresponding compressed videos as training samples. Data augmentation (i.e., rotation or flip) is further used to better exploit those training samples. We train all models using Adam optimizer (Kingma and Ba 2014) with β1 = 0.9, β2 = 0.999 and ϵ = 10 8. Learning rate is initially set to 10 4 and retained throughout training.