Slouch Tracker v1.1.0

What's new

• First-run onboarding wizard — on a fresh install the app now walks you through setup: pick your webcam (live preview with camera and person-detection status), then capture 5 good-posture frames, 5 bad-posture frames, and optionally 3 away frames. Skippable at any step; re-run it anytime via Settings → Run Setup Again.
• Camera selection — the capture device is now configurable and persisted (previously the first system camera was always used).
• Faster time-to-detection — training now starts automatically once the dataset has 3 frames per class (was effectively unbounded before; the wizard's captures are enough to produce a working model immediately).
• UI cleanup — model result cards no longer show "CV skipped" placeholders after automatic training runs; config sections are collapsed into disclosures.

Install

Download Slouch.Tracker_1.1.0_x64-setup.exe, verify against SHA256SUMS.txt, and run. Windows 10/11 x64 with a DirectX 12 GPU required. The NLF-L pose model (~245 MB) downloads automatically on first launch.

Slouch Tracker v1.0.0

First public release.

Slouch Tracker is a privacy-first webcam posture tracker for Windows. It watches your webcam, estimates your body pose in real time, and warns you when you start slouching. Everything runs on-device in native Rust: no cloud services, no accounts, no telemetry. The only network access the app ever makes is a one-time ~245 MB pose-model download on first launch.

Highlights

• Real-time posture detection: RTMDet-nano person detection plus NLF-L pose estimation (17 keypoints and 3D depth) on DirectML via native ONNX Runtime
• Train your own classifier in-app on your own labeled frames: 6 classifier types, 12 feature types, optional k-fold cross-validation with reported metrics
• Fast data collection with in-app keys and global hotkeys (Ctrl+Win+G/B/A) that work while the app is unfocused
• Privacy mode: the live preview is obscured while detection keeps running
• Minimize to tray, pause/resume from the tray or the window, optional start at login
• Local SQLite storage; dataset export and import as portable .slouchpack archives

Install

1. Download Slouch.Tracker_1.0.0_x64-setup.exe below and run it.
2. Windows SmartScreen will warn because the build is unsigned: click More info, then Run anyway.
3. On first launch the app downloads the pose model (~245 MB) once, verified against a pinned SHA-256. For machines without internet access, see the fully-offline installation steps in the README.

Requirements: Windows 10 or 11 (x64) with a DirectX 12-capable GPU.

Licensing

The source code is MIT. The NLF-L pose model (by István Sárándi, redistributed with the author's permission) is licensed for non-commercial use, so the application is for non-commercial scientific research, non-commercial education, or non-commercial artistic use cases only. Details in THIRD-PARTY-NOTICES.md.

Pose model (NLF-L crop, fp16 ONNX)

NLF-L crop pose model, exported to ONNX with an added backbone_feats output and cast to fp16. Downloaded automatically by Slouch Tracker on first launch.

SHA-256: 33bd300cd5a65681a5d671debd82a63f842c7420443cd9bb7424ca7aef82cca8
Size: 244,722,283 bytes

This is a conversion of the NLF-L model by Istvan Sarandi (https://github.com/isarandi/nlf), hosted with the author's written permission. The weights are for non-commercial use only; the Slouch Tracker application is for non-commercial scientific research, non-commercial education, or non-commercial artistic use cases only. Please credit the NLF paper and repository in derivative work.