African Actions Annotated Videos for AI (Health Domain)

African Actions Annotated Videos for AI (Health Domain)

Overview

African Actions is a multimodal, egocentric (first-person) video dataset of everyday activities performed in Nigeria. It was created to address two gaps in the data used to train modern AI: most human-activity datasets capture Western environments, and many are released under licenses that limit real-world use. African Actions brings African objects, tools, environments, languages, and ways of doing things into the data used to train embodied, agentic, and multimodal AI.

In this dataset, each video shows an action in the health domain, e.g. a healthcare procedure, recorded from the point of view of the person performing it, and is paired with fine-grained, time-stamped annotations describing what is happening.

Dataset statistics

• Videos: 10 egocentric MP4 recordings

• Annotated segments: 1,666 time-stamped action segments

• Distinct action–object units: 652 (the fundamental unit of the dataset)

• Distinct actions: 114 | objects: 160 | tools: 37 | secondary objects: 31

• Everyday activities covered: 8 distinct tasks, all in the "Health" domain.

• Indigenous-language descriptions (per segment): Igbo — 1,299; Hausa — 178; Yoruba — 164

• Natural-language coverage: 100% of segments have an English description; 99% have an indigenous-language description (the remaining are deliberate "No action" segments documenting moments of inactivity, which by design carry no translation)

What's in the archive

naijaactions-dataset/
├── videos/            MP4 files — egocentric recordings, one per video
└── annotations/       JSON files — one .json per video

The videos/ folder and the annotations/ folder pair one-to-one: every video has exactly one annotation file, and every annotation file points to a real video via its video_filename. The task_id → task name + domain mapping is published in the Tasks section of this dataset card, not as a separate file inside the archive.

Annotation schema

Each annotations/.json file looks like:

{
  "video_id": "na105_navr10",
  "video_filename": "NA105_NAVR10.mp4",
  "segments": [
    {
      "start": 5.404,
      "end": 11.064,
      "action": "pulling",
      "object": "clothes",
      "tool": "",
      "secondary_object": "",
      "nld_en": "The task performer who is a pharmacist is pulling the patient's clothes up to see the wound using hands.",
      "language": "Igbo",
      "nld_indigenous": "Onye na-agba ọgwụ na-ekpolite uwe onye ọrịa elu iji hụ ọnya ahụ"
    },
...

tool, secondary_object, language, and nld_indigenous may be empty strings when not applicable.

The action–object unit

The fundamental building block of the dataset is the action–object unit. This reflects a core design belief: the value of activity data lies not just in what action is performed, but in how it is performed — and "how" is shaped by the object involved, the tool used, and the environment. For example, cutting changes significantly when cutting a potato versus a plantain; the object influences the motion, technique, and even the tool selected. Capturing these fine-grained units lets models learn the authentic dynamics of human activity across diverse African settings.

Video filename structure

Each videos/*.mp4 file is named NA_NAVR_P.mp4, for example NA001_NAVR09_P1.mp4. The three components are:

• NA — three-digit task identifier (e.g. NA001). Every distinct task ID present in the dataset has exactly one row in the Tasks section of this dataset card, with the task's human-readable name.

• NAVR — the identifier of the on-the-ground recorder who filmed the session. Multiple recorders may cover the same task; the recorder ID lets you track per-recorder coverage and stratify by recording style without revealing the recorder's identity.

• P — sequential part number when a single task recording was split across multiple files (e.g. for very long sessions, or when the recorder paused between sub-activities). Parts of the same (task_id, recorder_id) are contiguous in time within the original recording session.

The corresponding annotations/.json file uses the same identifier with lowercase letters (e.g. na001_navr09_p1.json for NA001_NAVR09_P1.mp4). The video_filename field inside each annotation file always carries the canonical uppercase MP4 filename.

Languages

Natural-language descriptions are provided in English plus one of three major Nigerian languages — Igbo, Hausa, or Yoruba — chosen to match the linguistic and cultural context of each activity.

Activity domains

The full dataset uses 8 formal domain labels in its per-task index: Food/Kitchen, Art & Craft, Health, Construction & Fabrication, Compound & House Chores, Agriculture, Household, Fashion.

The activities themselves span everyday Nigerian life. The domain for this dataset is

• Healthcare & diagnostics — malaria/typhoid/HIV/pregnancy testing, setting an IV drip, herbal medicine preparation, bone setting (Health).

How the data was collected

Team. The on-the-ground team for this release comprised 10 recorders (identified in filenames as NAVR01–NAVR10), 17 trained annotators (identified internally as NAAN), 28 quality reviewers (some of whom are also annotators, wearing a second hat during the QC phase), plus a small group of native-speaker language consultants for Igbo, Hausa, and Yoruba who filled or verified every indigenous-language description in the curated release. A local coordinator, a welfare officer, and technical-support staff rounded out the project team.

Data collection followed a community-centred, human-in-the-loop methodology:

• Recording. Trained video facilitators recorded task performers (community members with deep knowledge of an activity) as they carried out a task naturally from start to finish. Separating the recorder from the performer responds to real-world conditions (familiarity with devices, age, comfort, cultural norms) and lets performers focus entirely on the activity.

• Egocentric viewpoint. Facilitators used a camera position that approximates a first-person view, capturing how the activity looks from the performer's own perspective. No scripted voice narration was required.

• Recording review. Each clip was checked for egocentric framing, clarity (sharpness/lighting), continuous capture of the activity, technical standards, and the absence of unintended content (e.g. bystander faces) before moving on.

• Annotation. Annotators labelled each pre-segmented video on a bespoke annotation platform across three layers: (1) action–object pair, (2) tool used, and (3) a natural-language description in English and an indigenous language.

• Annotation review & quality assurance. Separate reviewers validated annotations against detailed guidelines for internal consistency, linguistic accuracy, and alignment with the annotation layers, followed by a dedicated quality-assurance pass. Language consultants (native speakers) filled and verified the indigenous-language fields to ensure culturally accurate terms.

Quality control

The dataset passed through multiple independent review stages — recording review, annotation review, quality assurance, and a final language/oversight review — before curation. During curation, recordings that were rejected, duplicated, or failed quality checks were excluded, and a small number of segments and videos were removed via blacklists; the curated release in this archive contains only videos that pass these checks and have complete annotations.

Concretely, of the segments in scope for QC, 98.7% received at least one completed QC review. 1.7% of those reviews flagged a hard error (typically a label/description mismatch or a duplicate segment); 6.8% flagged a soft warning (style or vocabulary issues that did not block shipping). A dedicated correction phase subsequently revisited segments that had been flagged, applying fixes to roughly 28% of the input population; the corrected, re-reviewed segments are the ones included in this release. On top of these reviewer checks, an automated quality-report pipeline applied additional hard-gate filters (translation-verification status, language-allowlist enforcement, vocabulary integrity, NLD content presence, segment duration sanity, and video-level dropout-ratio thresholds) before the final release was packaged.

Consent audio. Verbal consent from each task performer was captured as a separate audio recording (one MP3 per performer/recorder pair) and is retained by the project team under access controls. These consent recordings are not bundled in this archive to protect performer privacy; they can be furnished to Mozilla or to a delegated reviewer on request for audit purposes.

Known characteristics and limitations

• Egocentric is a spectrum. Most videos approximate a first-person view but are not all perfectly head-mounted; some are captured close to the performer's eye level by a facilitator.

• Natural camera motion. Some shakiness from the activity itself is present and expected; sustained blur was screened out during review.

• Audio is secondary. Audio captures the ambient/environmental sound of each activity rather than clean speech; it is not intended for speech tasks.

• Short segments. Segments are fine-grained (median ~3 seconds), reflecting the action–object granularity of the annotations.

• Translation coverage. 2.1% of segments do not carry an indigenous-language description; these are the deliberate "No action" moments which by design require no translation. Every other segment ships with both English and an indigenous (Igbo, Hausa, or Yoruba) description that has been verified by a translation consultant.

• Sampling bias. The dataset reflects activities as practised by trained task performers recruited through community partners in Nigeria. It is not a randomly-sampled cross-section of Nigerian life and was not designed to be statistically representative of the population. Activity coverage is uneven: the Food/Kitchen domain alone holds 38% of tasks, while domains like Agriculture, Household, and Fashion are represented by only a handful of tasks each. Users building models intended to generalise beyond this dataset should evaluate carefully whether the captured distribution matches their target context.

• Environmental bias. Recordings were made in real working settings (kitchens, workshops, clinics, fields), not in laboratory studios. Lighting, background, and ambient noise vary across sessions. Some recording locations recur (multiple tasks were filmed in shared facilities such as the project's Aries office and Graceville, Nkwagu), which means tasks recorded in those locations share visual context that does not generalise to other places.

• No predefined train/validation/test splits. The archive is shipped as a single curated pool. Researchers are expected to construct their own splits according to their use case. Useful split keys include video_id (to avoid the same video appearing on both sides), NA (to evaluate task-level generalisation), and NAVR (to evaluate cross-recorder generalisation).

Tasks

The dataset spans 8 distinct tasks, identified by the NA prefix in each video filename. The table below maps every task ID present in the release to its human-readable name and domain.

Creators

The dataset was produced by Lanfrica Labs (the organisation behind the NaijaVoices initiative), with implementation partner Global KANAC, and was recorded and annotated on the ground in Nigeria by trained community members.