Skip to content

Latest commit

 

History

History
24 lines (18 loc) · 1.58 KB

README.md

File metadata and controls

24 lines (18 loc) · 1.58 KB
Raw

GSTVQA

Code for 'Learning Generalized Spatial-Temporal Deep Feature Representation for No-Reference Video Quality Assessment'. The code is partially borrowed from VSFA. image

Environment

  • Python 3.6.7
  • Pytorch 1.6.0 Cuda V9.0.176 Cudnn 7.4.1

Running

  • Download the pre-extracted multi-scale VGG features of each dataset from BaiduYun, Extraction code: gstv. Then put the features in the path: "./GSTVQA/TCSVT_Release/GVQA_Release/VGG16_mean_std_features/".

  • Train:
    python ./GSTVQA/TCSVT_Release/GVQA_Release/GVQA_Cross/main.py --TrainIndex=1 (TrainIndex=1：using the CVD2014 datase as source dataset; 2: LIVE-Qua; 3: LIVE-VQC; 4: KoNviD）

  • Test:
    python ./GSTVQA/TCSVT_Release/GVQA_Release/GVQA_Cross/cross_test.py --TrainIndex=1 （TrainIndex=1：using the CVD2014 datase as source dataset; 2: LIVE-Qua; 3: LIVE-VQC; 4: KoNviD）

Details

  • The model trained on each above four datasets has been provided in "./GSTVQA/TCSVT_Release/GVQA_Release/GVQA_Cross/models/"
  • The code for VGG-based feature extraction is available at: https://mega.nz/file/LXhnETyD#M6vI5M9QqStFsEXCeiMdJ8BWRrLxvRbkZ1rqQQzoVuc
  • In the intra-dataset setting, it should be noted that we use 80% of the data for training and the rest 20% data for testing. We haven't used the 20% data for the best epoch selection to avoid testing data leaky, instead, the last epoch is used for performance validation.