Video Summary - Visão Computacional na Beleza e Mídia - Morethanreal + Mantis - VEM AI [Temporada #11

Overview of the Episode (Tech Focus: Computer Vision + AI for Personalization and Video Intelligence)

The episode discusses how Brazilian/Latin American companies use computer vision and AI for two main application types:

Beauty/dermocosmetics + fashion personalization
- Virtual try-on
- Skin/hair analysis
Video intelligence / video processing
- AI pipelines that extract structured data from videos so it can be reused (e.g., highlights, brand exposure, detection tasks)

Case Study 1: MorenReal (Beauty + Dermocosmetics, Expanding into Fashion)

Product concept

MorenReal offers a platform for hyper-personalized consumer journeys in beauty/dermocosmetics and now fashion.

Computer vision capabilities

Analyze skin and hair from the user’s camera input
Measure/estimate pores, blemishes, and signs like expression lines
Drive personalized product recommendations based on the analysis
Enable virtual experimentation—users effectively “try makeup virtually” by placing their face in front of the camera

Virtual try-on (AR) mechanics

Uses real-time face tracking (points/contours for eyes and mouth)
Builds real-time 3D geometry of the face
Renders hyper-realistic materials with lighting/texture so the virtual product looks convincing

Mentioned challenge

Handling imperfect/bad camera inputs (e.g., crooked photos, poor lighting) while still identifying relevant facial features.

Business impact / performance claims

Reported conversion improvements, including “more than doubling” conversion rate for a sunscreen brand case
Example metric: reduced required site visits/sign-ups from about 100 people to <40 (later described as ~35 people to drive purchases)

Case Study 2: Mantis AI (Video Intelligence Layer with AI Pipelines)

Core product idea

An “agnostic video processing layer” that extracts useful data from video and repurposes it for customer-specific use cases.

What they extract (baseline pipeline)

A broad video extraction network with ~76 classes, including:

Segmentation by subject
Camera cut tagging
Scene/chapter detection
Audio event detection (e.g., who spoke, studio vs recorded sound)
Object and on-screen text detection

Orchestration + customization

A proprietary service orchestrator adds extra pipelines based on client needs.

Highlight detection (sports / live broadcasting)

Goal: detect editorially relevant moments (e.g., football highlights) and generate highlight reels
Live workflow: models run in the background while footage is airing, producing suggestions so editors don’t manually watch for hours
Outcome: editors receive pre-selected/structured clips (e.g., first half highlights prepared for social media)

Brand exposure analytics (advertising / sponsorship measurement)

“Brand exposure pipeline”: processes videos frame-by-frame to detect:
- brands
- their positions
- a heat map of brand presence on screen
Reports produced include:
- where brands appear (e.g., center vs corners)
- coordinates in the frame
- size/scale ratings (large/small/medium)
- approximate exposure duration
Business value / use cases:
- Sponsors and rights holders evaluate ROI and adjust assets (e.g., signage placement)
- Negotiation/accountability: exposure minutes and visibility patterns can be compared across brands
Notable anecdote: a brand (Mitsubishi) unexpectedly received higher exposure due to timing with an important highlight moment

Vision Language Models (VLMs): Explained and Positioned as Enabling Tech

Definition

VLM = Vision Language Model—a video-capable analog of LLMs.

Why they matter

VLMs can describe what happens in video over time
They produce searchable/usable structured data (potentially vectorized)

Use case analogy

In security/warehouse monitoring, an operator could ask:

“Did any incident happen today?” / “What time and where?”
“Show me the image,” pulling from multiple cameras

Cost/feasibility trend

Operating VLMs has become cheaper, making prototyping easier.

Shift: From “Object Detection + Training” to “General Models + Prompting”

The discussion contrasts two approaches:

Traditional AI: train models for specific object classes (limited domain flexibility)
VLM-driven approaches: more flexible—can respond to new prompts/use cases without the same retraining burden

Trade-off noted

Historically, VLMs struggled with tasks requiring precise bounding boxes, though newer models reportedly improved.

Agentic Development + Orchestration for Faster Iteration (Product + Engineering Workflow)

Core emphasis

Agentic systems are presented as a “point of no return” for productivity.

Examples of agent use

Developers interact via Slack with named agents that:
- test/execute code
- validate outputs
- increase throughput
Research teams repeatedly test:
- prompt combinations
- model combinations
- fine-tuning strategies
- new ideas beyond the original model’s primary purpose

Automating data annotation

Because labeling is expensive, they use LLMs/VLMs to pre-annotate frames
Humans then correct rather than label from scratch
Reported labeling time reduction: from weeks/months to ~2–3 days (for hundreds of frames)

System-Building Philosophy: Multiple Models + Decomposition

They argue it’s not “one model does everything.” Instead, build pipelines that combine strengths of different models, for example:

one step lists brands
another determines context/description
another extracts coordinates/placement

This is described as more accurate and controllable than one monolithic model.

Team/Leadership and Architecture Implications

Mindset change

Technology teams become enablers, not bottlenecks.

Software development impact

Roadmaps shift from months (2–3 months) to days via agents and faster prototyping.

Earlier validation

Build functional end-to-end prototypes sooner to reduce rework and improve alignment before large engineering investment.

Hiring/skills emphasis

“Agent management” doesn’t remove the need for core engineering skills:
- critical thinking
- systems architecture understanding
- evaluating whether agent-generated code is sensible

Main Speakers / Sources

Jomar Souza Silva — Developer Relationship Manager, NVIDIA Latin America
Marcos Trinca — CEO & Co-Founder, MorenReal
Mateus Icari — Co-founder & AI Officer, Mantis AI

Visão Computacional na Beleza e Mídia - Morethanreal + Mantis - VEM AI [Temporada #11 - Ep. 03]

Key takeaways