FEELER / v1.0
07
── A study of seven feelings

Read theunspoken.

A multi-stream neural architecture trained on 88,192 faces across six datasets. Five region streams attend to each other through transformer fusion to recognize seven core emotional states with state-of-the-art precision.

── The Catalogue
  • 01JOY
  • 02GRIEF
  • 03WRATH
  • 04AWE
  • 05DREAD
  • 06CONTEMPT
  • 07STILLNESS
Faces trained
88,192
across 6 datasets
Region streams
05
face / eyes / mouth / cheeks / forehead
Output dimensions
04
emotion / valence / arousal / intensity
Architecture
ResNet18 + Transformer
multi-stream fusion

How a face
becomes a feeling.

01
Detect

MTCNN locates the face. MediaPipe maps 468 landmarks. Five regions are isolated with surgical precision: the whole face, the eyes, the mouth, the cheeks, the forehead.

02
Encode

A pretrained ResNet18 reads the face. Four region CNNs read their patches. Each stream returns a vector of features — a numerical fingerprint of what is there.

03
Attend

A two-layer transformer lets the regions look at each other. Wide eyes ask the mouth what it means. The mouth answers. A single state emerges. Seven emotions. One label. One number.

joy · grief · wrath · awe · dread · contempt · stillness · joy · grief · wrath · awe · dread · contempt · stillness · joy · grief · wrath · awe · dread · contempt · stillness · joy · grief · wrath · awe · dread · contempt · stillness · joy · grief · wrath · awe · dread · contempt · stillness · joy · grief · wrath · awe · dread · contempt · stillness ·