MultimodalVideoModerator<T>
Models & Types · AiDotNet.Safety.Video
Comprehensive video moderator that combines frame-level content classification, temporal deepfake detection, and optional audio track analysis.
Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → VideoSafetyModuleBase<T> → MultimodalVideoModerator<T>
Implements: IGeneratedParameterRegistrar<T>
For Beginners
This is the “all-in-one” video safety checker. It looks at individual frames for harmful images, checks whether the video flows naturally between frames (deepfakes often don’t), and analyzes scene transitions to find where content might have been spliced in.
How It Works
Orchestrates multiple detection strategies for complete video safety analysis:
- Frame sampling with ensemble image classifiers (NSFW, violence, hate symbols)
- Temporal consistency analysis for deepfake detection
- Scene transition analysis for detecting spliced/manipulated segments
- Motion analysis for detecting unnatural movement patterns
References:
- Efficient video understanding via multi-scale temporal sampling (CVPR 2024)
- Spatio-temporal consistency for video deepfake detection (2025)
- Video content moderation at scale (Meta, 2024)
- VideoGuard: Multimodal video safety with reasoning-based instruction hierarchy (2025)
Constructors
| Constructor | Summary |
|---|---|
MultimodalVideoModerator(double deepfakeThreshold, double sceneChangeThreshold, double nsfwThreshold, double violenceThreshold) |
Initializes a new multimodal video moderator. |
Properties
| Property | Summary |
|---|---|
int LastSampledFrameCount { get; set; } |
Number of distinct frames the last EvaluateVideo call actually sampled. |
string ModuleName { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<SafetyFinding> EvaluateVideo(IReadOnlyList<Tensor<T>> frames, double frameRate) |
Fields
| Field | Summary |
|---|---|
int TaxonomyFrameBudget |
Number of image frames sampled per video, matching the paper’s annotation budget: “14 image frames, 1 thumbnail, and text metadata” were fed to the model for each of 19,422 videos. |