DeepFace
title: deepface type: language created: 2026-06-22 last_updated: 2026-06-22 related: ["radar/languages/Ultralytics", "radar/tools/FireRedOpenStoryline", "radar/tools/EditMind"] sources: ["https://github.com/serengil/deepface"] radar_quadrant: Languages & Frameworks radar_ring: Assess radar_position: inner
deepface is a Python library for face recognition, verification, and analysis. It wraps multiple underlying detection backends (OpenCV, RetinaFace, MTCNN, Dlib, YOLO, MediaPipe) behind a unified high-level API. Core functions: verify (same person check), find (search a database), analyze (age, gender, emotion, race), represent (embedding extraction), stream (live or file-based frame processing). MIT license, 23k stars, v0.0.100 released May 2026.
deepface's stream() function accepts a file path in place of a webcam index (parameter typed Any). Frame-level face recognition is performed sequentially. However, stream() returns None -- it displays results in a window rather than returning structured data. For pipeline use, the find() function is more suitable: it accepts an image or frame, queries a pre-built face database, and returns a list of DataFrames or dicts depending on the batched parameter.
No function returns media timestamps or frame numbers. The same wrapper pattern as radar/languages/Ultralytics applies: track frame index during iteration, convert to seconds via fps.
| Criterion | Coverage |
|---|---|
| Raw video processing | Partial -- stream() accepts file path but underdocumented; designed for webcam |
| Timestamps from user input | No -- no media timestamp in output |
| Face recognition | Yes -- verify, find, cluster, represent |
| Object detection | No -- face-only; no general object detection |
| Video editing | No |
| Agentic | No -- cloud MCP endpoint exists; OSS library has none |
| Pipeline composable | Partial -- find() returns DataFrames; stream() returns None |
| User friendly | High -- single-line API, no ML expertise required per documentation |
deepface fills the face identity layer that radar/languages/Ultralytics does not cover. ultralytics detects that a person is present in a frame; deepface identifies who that person is by matching against a reference database. Used together, a pipeline can locate frames containing a specific named individual and produce timestamped segment lists for downstream editing via radar/tools/FireRedOpenStoryline.
Placed in Languages & Frameworks / Assess / inner.
deepface is the most accessible face recognition library available: MIT license, 23k stars, single-line API calls, no GPU required for basic use, Docker and REST API supported. The inner Assess position reflects that the face identity use case is immediately relevant to any video discovery workflow and the library is straightforward to evaluate.
The ring stays at Assess rather than Trial because video file processing via stream() is underdocumented (return type is None, limiting pipeline composability), and the pre-v1.0 semver (v0.0.100) signals that the API surface is still considered unstable by the maintainer. The cloud MCP endpoint suggests a managed offering is emerging but is not part of the OSS library.
Trial gate: confirmed use of deepface find() or frame-level face matching against a reference database in a production video pipeline, with results consumed programmatically by a downstream tool.
