The dataset consists of a CSV file (139.4 kB) containing video streaming analytics metrics including resolution and bitrate data. This work was conducted by a collaboration of four research institutions: Simula Research Laboratory, Fraunhofer FOKUS, IHP GmbH, and Humboldt University of Berlin.

The 5GENESIS project (Grant agreement ID: 815178) was funded by the European Commission and ran from July 2018 to December 2021, focusing on 5G network testing and experimentation.

The dataset includes source videos encoded with the visually lossless ProRes 422 (HQ) codec and processed videos (PVS) encoded with ffmpeg 5.1.3 using the libx265 codec (HEVC) with CRF 20 and yuv420 chroma subsampling. Additionally, the repository contains recorded head rotation data (pitch, yaw, roll format), subject demographics, and questionnaire responses from subjective testing.

A subjective study was conducted with 30 participants to evaluate immersion, visual comfort, and exploration behavior for both stereoscopic and non-stereoscopic 360° videos. Participants watched HEVC encoded versions and rated the videos regarding presence, visual comfort, and quality. The key finding indicates that non-stereoscopic video viewing leads to slightly better presence, visual comfort, and quality ratings compared to stereoscopic variants, as stereoscopic versions exhibited visual artifacts that potentially degraded quality. Exploration behavior remained similar across both formats.

This work was funded by the Deutsche Forschungsgemeinschaft (DFG) as part of the ECoClass-VR project (DFG-444697733).

The dataset contains head rotation data and Simulator Sickness Questionnaire (SSQ) results from 48 participants who watched 20 different omnidirectional videos, each 30 seconds long, using the HTC Vive HMD. The research hypothesis was that since the SSQ was not originally designed for 360° video-related studies, it might be simplified while maintaining validity for assessing simulator sickness in this specific context.

Using Principal Component Analysis (PCA) across data from six previous studies, the researchers identified that a reduced questionnaire with only 9 out of 16 questions (less than 44% of the original) yields the best agreement with the complete SSQ. Exploratory Factor Analysis (EFA) confirmed that these nine symptom-related attributes sufficiently represent three key dimensions: Uneasiness, Visual Discomfort, and Loss of Balance. The study also compared the original SSQ with the Virtual Reality Sickness Questionnaire (VRSQ) and Cybersickness Questionnaire (CSQ).

The GitHub repository provides tools for evaluating both the SSQ and head rotation data, enabling researchers to apply the simplified questionnaire methodology to their own 360° video studies. The dataset is also available on Zenodo (DOI: 10.5281/zenodo.4472672) with open access under Creative Commons Attribution 4.0 International License.

The dataset comprises three subjective test rounds with a total of 191 stimuli (64 stimuli in tests 1 and 2, 63 stimuli in test 3) at resolutions of 4K, 6K, and 8K using real-world streaming bitrates. Testing was conducted using HTC Vive for the initial test and HTC Vive Pro for subsequent tests, with quality assessments collected using the 5-point Absolute Category Rating (ACR) scale.

The dataset provides comprehensive information for each stimulus including source video links, subjective quality scores, video metadata, head rotation tracking data, and Simulator Sickness Questionnaire (SSQ) responses per stimulus and participant. This rich multimodal data enables research on the relationship between user behavior (head movements), physiological responses (simulator sickness), and perceived video quality in 360-degree streaming scenarios.

Multiple quality metrics were evaluated including VMAF, PSNR, SSIM, ADM2, WS-PSNR, and WS-SSIM. The key finding showed that VMAF demonstrated the highest correlation with subjective scores among all tested metrics. A center-cropped VMAF variant achieved comparable performance while enabling significantly faster computation, making it more practical for real-time streaming applications.

The research developed two lightweight quality prediction models: a bitstream metadata-based approach using only encoding parameters, and a hybrid no-reference model utilizing bitrate, resolution, and pixel data. Both models achieved performance levels similar to full-reference approaches while requiring substantially less computational resources, making them suitable for integration into adaptive streaming systems.

The dataset is hosted on Zenodo (DOI: 10.5281/zenodo.4090961) under Creative Commons Attribution 4.0 International License.

The dataset comprises 135 images generated using five different AI text-to-image generators (including DALL-E-2, Midjourney, and Craiyon) based on 27 different text prompts. Some prompts were derived from the DrawBench benchmark, ensuring diversity in scene complexity and description specificity. The generated images were combined with real photographs for comparison in subjective evaluation studies.

A companion paper at QoMEX 2023 (DOI: 10.1109/QoMEX58391.2023.10178486) extends this analysis by evaluating quality and appeal through crowdsourcing tests and correlating subjective ratings with objective quality assessment models.

The repository is structured with an AVrateVoyager folder containing the modified software for collecting mouse movement data, and a dataset folder with the actual mouse movement recordings. This behavioral interaction data provides insights into how users navigate and interact with video content during quality assessment tasks, particularly in different testing environments.

The dataset contains 220 video recordings (MOV format, 7680x3840 resolution) and 200 single-channel audio recordings (WAV format, 48000 Hz) of 20 native German speakers arranged in a circular classroom configuration. Each speaker read ten different stories, plus recordings of speakers in silence. The dataset includes 65 pre-rendered sample videos (MP4 format) and example subjective data from participants.

Two distinct audiovisual scenarios are provided: 360° omnidirectional video and computer-generated imagery (CGI). The implementations support both binaural audio synthesis using the Virtual Acoustics auralization framework and diotic presentations. The CGI version is built in Unity and supports HTC Vive controller interaction, while both scenarios enable investigation of how acoustic representation (diotic vs. binaural) and visual representation (360° video vs. CGI) impact cognitive performance and audiovisual scene analysis.

The companion repository AVT-ECoClass-VR-dataset (DOI: 10.5281/zenodo.14019040) provides subjective experiment data from three studies with 94 total participants (36, 24, and 34 respectively), including head/controller tracking with timestamps, speaker-story mappings, and questionnaire responses covering simulator sickness (SSQ), presence (IPQ), workload (NASA RTLX), listening effort, and noise sensitivity assessments. The subjective results were published in Frontiers in Psychology, Volume 16, 2025 (DOI: 10.3389/fpsyg.2025.1520630).

This was a collaborative work between TU Ilmenau, RWTH Aachen University, and RPTU Kaiserslautern.

The dataset contains 175 total source videos with 34 subjects per lighting scenario. The structure includes 4 distinct lighting scenarios plus one looped scenario combining all lighting conditions, enabling research on how lighting affects perceived quality of face videos in high dynamic range format.

This work was supported by the Deutsche Forschungsgemeinschaft (DFG), the ILMETA project, and the PoQuMo8K project.

The dataset contains four subjective test datasets (test_1 through test_4) with videos encoded using three different video codecs: H.264, HEVC, and VP9. The resolutions of the compressed videos range from 360p to 2160p with frame rates varying from 15fps to 60fps. All source 4K contents use 60fps, maintaining 3840 × 2160 pixel resolution (UHD-1).

The database includes encoded video segments, source videos, subjective ratings presented as Mean Opinion Scores (MOS) with confidence intervals, objective quality scores, and additional metrics including BRISQUE, NIQE, and VMAF reports. This comprehensive collection enables research on codec performance comparison and quality model validation.

The database has been utilized in multiple QoMEX conference publications, particularly for ITU-T P.1204.3 video quality model evaluation and related video quality assessment research.

The dataset contains short video segments (8-10 seconds each) from various sources, encoded using three modern video codecs: HEVC/H.265, AV1, and VVC/H.266. Videos span multiple resolutions (360p to 2160p) and bitrates (100kbps to 15mbps), enabling comprehensive analysis of encoding parameter effects on perceived video quality and appeal.

The subjective methodology employed a unique three-part experimental protocol. First, participants rated the appeal of uncompressed UHD-1/4K source content. Second, they assessed the quality of encoded versions across various codec, resolution, and bitrate combinations. Third, participants re-rated the source content appeal, allowing investigation of the bidirectional relationship between appeal and quality perception. This approach enables research on how encoding-related degradations interact with inherent content appeal.

The database includes both subjective scores collected via the AVRateNG tool and objective quality predictions from multiple models including VMAF, SI/TI values, hybrid model (hyn0) scores, and no-reference (nofu) scores. The research examines how different encoding conditions impact perceived video quality and investigates whether source content appeal influences quality perception and vice versa.

This dataset is part of the broader AVT-VQDB-UHD-1 series from TU Ilmenau’s Audiovisual Technology Group, extending the original database with explicit appeal assessment alongside traditional quality evaluation.