DOVER

Free

Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives.

FreeFree tier

Inputs: video

Type

Open Source

Visit Website GitHub (40 stars)

About DOVER

DOVER (Disentangled Objective Video Quality Evaluator) is an open-source tool from the ICCV 2023 paper that separates video quality assessment into aesthetic and technical perspectives for user-generated content (UGC). Developed by VQAssessment, it provides official code, pretrained weights, and evaluation scripts. The full DOVER model achieves state-of-the-art performance on benchmarks like KoNViD-1k (PLCC 0.883), LIVE-VQC (0.854), and LSVQ (0.889). A lightweight variant, DOVER-Mobile, uses only 9.86M parameters and runs on CPU at 1.4s per video. The repository includes training code, ONNX conversion, and a demo for evaluating single videos or whole datasets. All weights are available on Hugging Face.

Key Features

Disentangled aesthetic and technical quality assessment

DOVER-Mobile lightweight variant with 9.86M parameters (5.7x fewer than DOVER)

Pretrained weights available on Hugging Face

Full training code and evaluation scripts included

ONNX conversion support for deployment

Demo tool for predicting quality scores on single videos or sets

Pseudo-labelled quality scores for Kinetics-400 and YFCC-100M datasets

Pros & Cons

Pros

State-of-the-art performance on KoNViD-1k, LIVE-VQC, LSVQ (PLCC 0.883, 0.854, 0.889)
Lightweight DOVER-Mobile runs on CPU at 1.4s per video, 5.4x fewer GFLOPs than DOVER
Open-source with permissive S-Lab License and easily installable via pip
Provides separate aesthetic and technical quality scores for deeper analysis
Includes pretrained weights, demo scripts, and training code for reproducibility

Cons

Full DOVER model requires GPU for practical inference (3.6s per video on CPU)
Primarily designed for user-generated content; performance on other video types may vary
Documentation as a research repo; less beginner-friendly than commercial tools
Limited community support beyond GitHub issues

Best For

Evaluating video quality of user-generated content (UGC)Benchmarking video quality models across multiple datasetsTraining custom video quality assessment models with provided codeResearch into disentangled quality perspectives (aesthetic vs. technical)Deploying lightweight quality assessment on edge devices (CPU)

FAQ

What is DOVER?

DOVER stands for Disentangled Objective Video Quality Evaluator. It is an ICCV 2023 paper and open-source tool that assesses video quality from both aesthetic and technical perspectives in user-generated content.

How can I install DOVER?

Clone the repository (git clone https://github.com/VQAssessment/DOVER.git), then run 'pip install -e .' inside the DOVER directory. Download pretrained weights as described in the README.

How do I switch to DOVER-Mobile?

Add '-o dover-mobile.yml' at the end of your training, testing, or validation scripts to use the lightweight mobile backbone.

What are the performance numbers for DOVER?

On benchmark datasets: KoNViD-1k PLCC 0.883, LIVE-VQC PLCC 0.854, LSVQ_test PLCC 0.889, LSVQ_1080p PLCC 0.830. DOVER-Mobile achieves 0.853, 0.835, 0.867, 0.802 respectively with faster CPU inference (1.4s per video).