Video Sources

Acquisition of a representative dataset involves selecting a large number of videos with suitable technical parameters (length, license, etc.), then making representative subsamples using clustering and feature selection methods; after that, we plan to apply distortions to the resulting videos and the resulting representative set becomes ready for subjective comparison.

Requirements to Video Sources

• Only videos with CC-BY or CC-0 licenses (re-distribution rights).
• Duration: at least 10 seconds.
• Frame rate: at least 10 FPS.

Selected Video Sources

• Vimeo
• FineVideo
• YouTube

Video Pre-processing

• No more than 2 slices are selected from the list of scenes detected using AutoShot.
• The number of scenes in each slice is minimal, so that the total length of the fragment is at least 15 seconds.
• Slices are selected to be equidistant from each other, the beginning, and the end of the video; final fragments are trimmed to 15 seconds in the center.
• Encoding (FFmpeg): -vcodec libx264 -pix_fmt yuv420p -crf 0 -preset medium -sn -dn
• Audio (if present): -acodec aac -audio_bitrate 256k
• Frames resized to 1080 pixels on the smaller side while keeping the aspect ratio. So the final resolution was 1920×1080 and 1080×1920

Feature Extraction

To properly select videos for representative subsampling we utilized various feature extraction methods.

General IQA/VQA Metrics

• Luminance histogram features: luminance quantiles, dark and bright pixel ratios, noise ratio, and entropy
• Hasler–Suesstrunk
• Std-Luminance
• CLIP-IQA+
• TOPIQ
• LAION Aesthetic
• PaQ-2-PiQ
• StableVQA

VQMT Metrics

• SI and TI
• Blurring
• Blocking

Semantic Features and Labeling

• YOLOv12-L
• Places365
• SigLIP-2
• InternVL3

Feature-Based Video Clustering

Clustering Techniques

• Each video is represented by a unified feature vector combining numerical quality metrics, categorical semantic descriptors, and InternVideo embeddings.
• Numerical features are summarized by the mean and standard deviation of frame-level quality metrics.
• Categorical features are constructed from annotations using occurrence or confidence-weighted counts, then log-scaled and ℓ2-normalized: \(x_c \leftarrow \dfrac{x_c}{\left\lVert \log(1+x_c)\right\rVert_2 + \varepsilon}\)
• Embeddings are normalized to unit length: \(x_e \leftarrow \dfrac{x_e}{\left\lVert x_e\right\rVert_2 + \varepsilon}\)
• All components are concatenated and globally ℓ2-normalized to obtain the final representation x.
• Clustering: k-means in FAISS with k = 1000 clusters and 50 training iterations, minimizing squared Euclidean distance in the normalized feature space.
• After clustering, each video is assigned to its nearest centroid; the video with the smallest distance to the centroid is selected as the representative of each cluster, yielding a diverse and balanced subset of videos.
• The result is a selection of 1000 videos for inference methods and comparison using a subjective quality rating system (QRS).

Distribution of cluster sizes after k-means clustering.