Main application interface showing file selection, model options, and confidence threshold controls
The TinyExplorer Detection App is a user-friendly graphical interface designed with developmental scientists in mind. This toolbox integrates state-of-the-art open-source face recognition algorithms into an easy-to-use software package, streamlining the process of analysing facial data.
- Simple graphical user interface for easy operation
- Integration of cutting-edge face recognition models
- Batch processing capabilities for efficient analysis of large datasets
- Customizable confidence thresholds for detection accuracy
All processing is performed entirely on your local machine. No images, videos, or detection results are ever uploaded to external servers. Your research data stays completely private and under your control.
- No cloud processing or data transmission
- No internet connection required after initial model download
- Models are downloaded once and stored locally
- Ideal for working with sensitive research data involving human subjects
Choose from multiple face detection models included in the app:
- YOLOv8n-face (Nano): fastest inference, smallest size (~2.7 MB); lower accuracy; ideal for real‑time or limited resources.
- YOLOv8m-face (Medium): balanced speed and accuracy (~27.3 MB); solid default for most tasks.
- YOLOv8l-face (Large): highest accuracy within v8 (~59.2 MB); slower inference; best for high precision.
- YOLOv11m-face (Medium): newer generation with improved accuracy/speed trade‑offs; good general‑purpose choice on modern hardware.
- YOLOv11l-face (Large): higher accuracy variant; increased compute and memory cost.
- YOLOv12l-face (Large): latest large model; highest accuracy and resource use; recommended for offline batch processing.
- RetinaFace: alternative architecture with facial landmarks; good speed/accuracy for feature localization. Note: available on Apple Silicon (arm64) macOS only.
Choose from multiple hand detection models included in the app:
- HandObject (100DOH baseline) - Hand_object_detector model by ddshan, trained on 100DOH dataset
- HandObject (100DOH TinyExplorer-tuned) - TinyExplored-Tuned version of the 100DOH hand_object_detector, targeted at detecting infant hands with ownership classification integrated
In order to use the 100DOH TinyExplorer-tuned model within the app, you must first gain access to the model via HuggingFace and provide an access token.
Find the link to the model here: https://huggingface.co/ThompsonC21/100DOH-TinyExplorer-Tuned-hand-detection
Find information on obtaining an access token here: https://huggingface.co/docs/hub/en/security-tokens
The Speech mode extracts timestamped transcripts from audio files and videos. It
supports OpenAI Whisper, Faster Whisper, and WhisperX. Results are written to a
timestamped folder as both <file>_transcript.txt and
<file>_transcript.csv, plus shared detections.csv and summary.csv files
for aggregate processing. The shared CSV uses common identity/source fields
alongside speech timestamps, text, language, and model information.
The ASR packages are included in newly bundled installers. For development,
install them in the active Python environment with pip install -r python/requirements-whisper.txt. Model weights are downloaded on first use and
remain local.
The app automatically downloads required model weights when needed.
Measured 2026-09-02 with the bundled Windows build on a clean Windows 11 Pro
VM (Azure Standard_B4s_v2: 4 vCPU, 16 GB RAM, Microsoft Defender active).
Transcription time for a 64-second spoken-audio clip, CPU only, warm caches:
| Model | Faster Whisper | OpenAI Whisper | WhisperX |
|---|---|---|---|
| tiny | 3.5 s | 8.2 s | 3.9 s |
| base | 6.6 s | 9.9 s | 6.0 s |
| small | 19.2 s | 23.1 s | 15.3 s |
| medium | 56.5 s | 63.3 s | 42.0 s |
| large-v2 | 97.9 s | 114.4 s | 73.4 s |
| large-v3 | 97.0 s | 114.9 s | 72.1 s |
| large-v3-turbo | 56.4 s | 200.2 s * | 51.1 s |
* OpenAI Whisper runs fp32 on CPU; with large-v3-turbo it fell into temperature-fallback re-decoding and hallucinated extra text — avoid that combination on CPU-only machines.
Reading the table: the default small runs at 3-4x realtime on all backends;
WhisperX is the fastest backend from medium up (batched inference) and even
large-v3 stays close to realtime with it. Warm model load adds roughly 1-10 s
per run (WhisperX 9-20 s, since it also initialises a voice-activity pipeline;
OpenAI Whisper large ~20 s).
Cold start — the very first use of a model — additionally downloads weights and warms caches: a few seconds for tiny/base, ~10 s for small, ~30-60 s for medium, and ~1-2.5 min for the large variants (~3 GB download each) on a fast connection; WhisperX's first-ever run adds a further one-off ~100 s of pipeline initialisation. During this phase the app can look idle — that is the "first use may download model weights" stage, not a hang.
Note: on a freshly installed Windows the app additionally requires the Microsoft Visual C++ Redistributable (x64) — without it the detection and speech backends fail to start. Installers built after 2026-09-01 detect and install it automatically; for older builds, install it manually from https://aka.ms/vs/17/release/vc_redist.x64.exe.
- YOLO face weights: https://github.com/cardiff-babylab/tinyexplorer-detection-app/releases/tag/v1.0.0-models (originally from https://github.com/akanametov/yolo-face)
- RetinaFace implementation: https://github.com/serengil/retinaface
- Hand detection weights: https://github.com/cardiff-babylab/tinyexplorer-detection-app/releases/tag/handobj-weights-v1 (HandObject / 100DOH Faster R‑CNN, originally from https://github.com/ddshan/hand_object_detector)
- Whisper (OpenAI): openai/whisper
- Faster Whisper: SYSTRAN/faster-whisper
- WhisperX: m-bain/whisperx
This toolbox addresses several key needs in developmental psychology research:
- Efficiency: Automates the time-consuming process of manual face detection in video and image data.
- Accessibility: Provides a user-friendly interface, making advanced face recognition technology accessible to researchers without extensive programming experience.
- Flexibility: Allows researchers to easily switch between different face recognition models to suit their specific research needs.
- Reproducibility: Ensures consistent application of face detection criteria across studies, enhancing research reproducibility.
This toolbox is actively developed by the Cardiff University BabyLab, a research group dedicated to exploring attentional and motor skills in young children and their impact on learning in everyday settings. We welcome contributions from the developmental psychology community to enhance and expand the capabilities of this toolbox.
If you have ideas for new features, improvements, or bug fixes, please feel free to:
- Submit a pull request
- Open an issue with your suggestion
- Contact us directly with your ideas
- Download the latest installer for your OS from the Releases page.
- Run the installer and launch the app following the instructions in the Getting started page.
If no release exists for your system or the installer doesn't work, you can build locally.
- Ensure Node.js and npm are installed.
- Install dependencies:
npm install - Start in development:
npm run start - Build a distributable package:
npm run build
- Basic Usage: https://cardiff-babylab.github.io/tinyexplorer-detection-app/getting-started
- Supported File Formats: https://cardiff-babylab.github.io/tinyexplorer-detection-app/main-features/#supported-file-formats
For more information or collaboration inquiries, please contact:
- Organization: Cardiff BabyLab, Cardiff University Centre for Human Developmental Science (CUCHDS)
- Address: 70 Park Place, Cardiff, CF10 3AT, UK
- Email: babylab@cardiff.ac.uk
- Phone: 029 2251 4800
- Website: cardiff-babylab.com
We look forward to seeing how this toolbox can support and advance your research in developmental psychology!
This work was supported by a James S. McDonnell Foundation (JSMF) Opportunity Award and a UKRI Future Leaders Fellowship (MR/X032922/1) awarded to HD.