Desempenho facial: 3 dicas essenciais para protocolos de captura de dados

Ao projetar protocolos de captura facial, há muitos aspectos que devem ser considerados para evitar a produção de dados de baixa qualidade e a fadiga dos participantes.

Fatores como...

quais poses você escolhe
como você sequencia as poses
como você explica/mostra as poses, etc.

...fazem uma enorme diferença no resultado de suas sessões.

Working with major game and tech companies to refine their facial performance capture pipelines, it’s quite obvious that people are recycling similar, dated protocols. Considering the present purpose of many captures, such old protocols are often clunky with illogical pose combinations, redundant expressions, and inefficient flow.

Há uma lista interminável de prós e contras no design de captura facial, mas aqui estão algumas dicas gerais:

1. Be use case-minded!

Are you defining machine learning training data for real-time applications? Are you gathering extreme poses for high-intensity fight scenes in film or games? Think about your end goal.

If your purpose is to capture facial data for an avatar product aimed toward co-working, you’d likely want to prioritize prosocial, collaborative, and natural facial expressions. Stop wasting time, energy, and production budget by bloating your session with every possible unsightly “scream” pose or hyper-compressed “lemon” face. Save those for Diablo e Planet of the Apes.

2. Design for logical flow.

Group similar expressions together, e.g. brow-based FACS poses (action units like – inner brow raiser, outer brow raiser, brow lowerer) and eye-based FACS poses (action units like – upper lid raiser, lid tightener, cheek raiser, eye closure, blink, wink) together.
Go from easy to difficult within each section. If you move from easy to difficult across ALL poses, you will end up forcing your user to leap from eyes to mouth to brows to jaw.
Grouping expressions strategically not only helps with user fatigue and comprehension, but it also opens up opportunity to order your poses in a way that allows you to describe and build off previous ones.

3. Make sure your example imagery and descriptions match the target pose and intended blendshapes.

Too many times, I see prompts like “raise brows without widening eyes” – yet the actor in the example shot is clearly widening their eyes. A large percentage of users will do as they see, not as they hear or read. So, don’t give conflicting instructions, and make sure you rigorously review the example poses!

📝 Mais uma dica: No vídeo, observe como eu indico a ação conhecida como elevação do queixo dizendo "empurre o lábio inferior para cima". É fácil ficar preso à descrição de poses com base em seus nomes formais, mas, ao usar descritores mais acessíveis, você pode aumentar a probabilidade de o participante acertar a pose.