Menu
Back to projects

EDITH Intake

  • Multimodal Research / Vision
  • Research

Signals need context.

Research in computer vision and sensor data for lower-friction intake tracking.

Implemented software research; the wearable hypothesis remains unvalidated.

View public research
Flowing dark ribbons lit with violet and cyan highlights.

Software to investigate.

The public source snapshot records working perception, depth and temporal software. Each foundation can be tested independently.

Implemented software

At that snapshot, live physical capture still required operator action. Capture tooling does not establish worn-device performance.

Read the software evidence
Perception & identity
DINOv2 visual embeddings, multiple reference prototypes and unknown-class rejection provide a software baseline for identity experiments.
Monocular depth
Depth Anything V2 Small runs locally on GPU and CPU. Calibration and quantity ablations remain research; physical portion measurement is unvalidated.
Temporal & IMU processing
SI-unit contracts, sliding windows and signal transforms feed a four-state event model. Debouncing, hard-negative cases and a separate drinking path are software foundations.
Capture & provenance tooling
Mobile capture tooling and CSV/JSONL import preserve SHA-256 provenance, normalize units and audit timing, gaps and non-monotonic timestamps.

Complementary signals.

Why combine signals? Each addresses a different weakness. This role map describes a research architecture; the full multimodal system is not established as an integrated wearable.

IMUImplemented software
Candidate-event detectionDoes not establish food identity.
RGBResearch / experiment
Identity and visual contextAn image alone does not establish absolute scale.
Monocular depthImplemented software
Dense relative geometryAbsolute metric scale still requires calibration.
Time-of-Flight / ToFSystem hypothesis
Potential metric geometry constraintPhysical integration remains a future validation step.
Temporal evidenceResearch / experiment
Change across observationsConsumed-versus-served reasoning remains a research question.
Read the research architecture

Evidence keeps its label.

A result keeps the scope of its evidence. Software, reference fixtures and simulations do not establish physical or wearable accuracy.

Power estimates remain simulated. Reference-fixture quantity results are separate from physical meal measurements.

Read the evidence boundary

  1. Synthetic fixture

    Controlled inputs exercise software behavior.

  2. Reference harness

    Repeatable comparisons test selected assumptions.

  3. Model inference

    Real model execution demonstrates its software path.

  4. Controlled physical measurement

    Ground-truth measurements must support physical claims.

  5. Wearable field evidence

    Worn-device performance needs its own validation.

A hypothesis to test.

Can a small wearable reduce nutrition-logging friction by combining complementary signals rather than depending on a single image?

System hypothesis · Future hardware validation

Event-triggered observations and heavier inference on a paired phone are proposed directions. Hardware integration, physical data collection and wearable field validation remain separate research steps.

The public research edition contains real software evidence. Its wearable hypothesis remains unvalidated.