---
license: mit
tags:
  - facial-expression
  - action-units
  - emotion-recognition
  - gaze-estimation
  - benchmark
pretty_name: py-feat benchmarks
configs:
  - config_name: accuracy
    data_files: accuracy.csv
  - config_name: throughput
    data_files: throughput.csv
---

# py-feat benchmarks

Live benchmark data for [py-feat](https://github.com/cosanlab/py-feat) and a
cross-tool comparison against **OpenFace 3.0**, **LibreFace**, and **PyAFAR**.
Powers the [py-feat live dashboard](https://py-feat.org). Updated by scheduled
benchmark runs.

## Files

| File | What |
|---|---|
| `accuracy.csv` | Tidy long table: one row per `(tool, dataset, modality, metric)`. Covers AU F1 (DISFA+), 7-class emotion (AffectNet-val, RAF-DB), valence/arousal CCC (AffectNet-val), and gaze angular error (Columbia). |
| `throughput.csv` | py-feat detector frames/sec across hardware × batch. |
| `cross_tool_methodology.md` | How every tool was run end-to-end, the feature matrix, dataset/protocol choices, and per-tool integration notes. |
| `competitors/*.json` | Raw per-tool result JSONs (provenance for every number in `accuracy.csv`). |

## `accuracy.csv` columns

`tool` · `dataset` · `modality` (au / emotion / valence_arousal / gaze) ·
`metric` · `value` · `n` (samples scored) · `notes`.

## Protocol in one paragraph

Every competitor runs **end-to-end as shipped** (its own face detector, its own
models) on **identical images/labels** frozen into shared manifests — never as a
py-feat dependency, each in its own isolated env. Held-out sets only (DISFA+,
not DISFA, since DISFA trains several competitors). Emotion is scored on the 7
shared classes; gaze I/O conventions are resolved identically and in every
tool's favor. See `cross_tool_methodology.md` for the full story, including what
it took to get each competitor to run.