AI Face Tracking & Smart Reframing for Vertical Video Conversion
A content repurposing platform needed to automatically convert horizontal (16:9) long-form videos into vertical (9:16) short-form clips while keeping speakers and subjects perfectly centered — without any manual cropping or keyframing.
Discutez de Votre Projet
Le DĂ©fi
Converting horizontal video to vertical format was one of the most tedious steps in short-form content production:
- Manually cropping and repositioning the frame for every clip was time-consuming
- Multi-person conversations required dynamic reframing as speakers changed
- Static center-crop cut off speakers who moved or sat off-center
- Traditional face detection was too slow for real-time reframing decisions across thousands of clips
- Different content types (interviews, solo vlogs, presentations) required different framing strategies
Notre Solution
We built an AI-powered face tracking and smart reframing engine that detects faces in video frames, tracks their movement, and dynamically adjusts the vertical crop region to keep the active subject centered.
Architecture
- Face Detection: YOLO-based face detection model optimized for speed
- Face Tracking: IoU-based frame-to-frame tracking with persistent subject IDs
- Reframing Engine: Dynamic crop region calculation based on face positions and movement
- Active Speaker Coupling: Integration with speaker detection to prioritize the person talking
- Rendering: FFmpeg crop filter chain with smooth pan transitions
Reframing Pipeline
- Face Detection - Run YOLO face detection across sampled frames
- Subject Tracking - Link face detections across frames using IoU-based tracking
- Speaker Priority - When coupled with active speaker detection, prioritize the talking subject
- Crop Calculation - Determine optimal 9:16 crop region based on primary subject position
- Smoothing - Apply easing to crop movement to avoid jarring jumps
- Rendering - FFmpeg applies the dynamic crop with smooth pan transitions
Key Features
- Multi-Subject Handling - Tracks multiple faces and determines the primary subject per segment
- Speaker-Aware Framing - Prioritizes the active speaker when integrated with speaker detection
- Smooth Transitions - Eased panning between subjects eliminates jarring cuts
- Content-Type Adaptation - Different framing strategies for solo, interview, and group content
- Batch Processing - Reframe hundreds of clips from a single long-form video
- No Manual Intervention - Fully automated from detection to final render
RĂ©sultats
Stack Technologique
caseStudyDetail.more Études de Cas
Découvrez plus de nos implémentations techniques
Planification et analyse de la performance des médias sociaux multiplateformes
Les créateurs de contenu produisant des dizaines de clips courts chaque semaine avaient besoin d'un système unifié de planification et d'analyse pour distribuer du contenu sur TikTok, YouTube Shorts et Instagram Reels à partir d'un tableau de bord unique — avec des aperçus pour optimiser la stratégie de publication.
Traduction de sous-titres multilingues pour la distribution mondiale de contenu
Les créateurs de contenu avec des audiences internationales avaient besoin d'étendre leur portée en traduisant les sous-titres vidéo dans plus de 30 langues tout en préservant l'audio original, permettant aux spectateurs du monde entier de consommer le contenu dans leur langue maternelle.
Questions fréquemment posées
MicrocosmWorks implemented a hybrid tracking approach that combines a lightweight face detector running every 5th frame with a KCF optical flow tracker for inter-frame predictions. When occlusion is detected via confidence score drops, the system maintains the last known trajectory with Kalman filtering and re-acquires the face within 200ms of it becoming visible again.
MicrocosmWorks built a saliency-weighted cropping algorithm that prioritizes detected faces, then text regions, then motion areas when determining the 9:16 crop window position. For multi-person scenes, the system uses a configurable priority ranking, defaulting to the active speaker or the largest face, with smooth interpolation between crop positions to avoid jarring shifts.
Yes, MicrocosmWorks implemented a fallback saliency detection mode that activates when no faces are present, using a combination of motion detection, visual attention modeling, and mouse cursor tracking for screen recordings. The system intelligently follows the most relevant content region even in purely visual or text-based footage.
MicrocosmWorks optimized the pipeline for batch workflows, achieving 8x real-time processing speed on a single NVIDIA T4 GPU, meaning a 10-minute video is reframed in approximately 75 seconds. The system supports parallel processing across multiple GPUs, scaling linearly for high-volume content operations.
MicrocosmWorks develops AI video reframing systems at rates of $25-$45/hr, with a full face tracking and smart reframing solution including model optimization, batch processing support, and API integration typically requiring 350-550 development hours. This investment eliminates the need for manual reframing editors, which typically cost $5-$15 per video.
PrĂŞt Ă Transformer Votre Entreprise ?
Discutons de la façon dont nous pouvons appliquer des solutions similaires à vos défis.