Skip to content

Datasets

Two sample datasets, each a Python dataset file plus a video collection, are available for training and testing custom VMAF models. Point the Python harness at the dataset file, and put the downloaded videos where the file's ref_dir and dis_dir say.

Dataset Dataset file Reference / distorted videos Access
Netflix Public Dataset NFLX_dataset_public.py 9 / 79 Public download, access granted on request
VQEG HD3 Dataset VQEGHD3_dataset.py 8 / 63 CDVL account, then convert to YUV420P

Use a dataset

The dataset files are consumed by the Python harness:

Netflix Public Dataset

The dataset is publicly available to the community for training, testing and verification of results. The videos are in YUV420P format and can be downloaded from this folder (access is granted on request).

File naming

  • Each file name has the form {content name}_{expert score}_{height in pixels}_{bitrate in Kbps}.yuv.
  • For example, BirdsInCage_85_720_1050.yuv is a decompressed video of the content BirdInCage, with an expert opinion score of 85 out of 100, compressed at 720p and 1050 Kbps.
  • The expert opinion score differs from the MOS offered in NFLX_dataset_public.py, which comes from a panel of non-expert subjects.

VQEG HD3 Dataset

The dataset file contains the file names of VQEG (Video Quality Experts Group) HD3 videos. The videos themselves are available from CDVL.

Download the videos

  1. Log in to CDVL.
  2. Choose "Advanced Search".
  3. Select "VQEG Subjective Tests" in the dataset dropdown and search for the keyword vqeghd3. The result is a single combined ZIP file of about 84 GB, "VQEG HDTV Test, vqeghd3", because the individual sequences are no longer available for download.
  4. Convert the videos to YUV420P format.

Selection

The dataset file includes src01 to src09 except src04, which overlaps with the Netflix Public Dataset. It includes hrc04, hrc07, hrc16, hrc17, hrc18, hrc19, hrc20 and hrc21, which are the distortion types most relevant to adaptive streaming.