Got questions about SentinelLabeler? Here are detailed answers regarding our architecture, hardware requirements, privacy, and workflows.
Never. We built SentinelLabeler on a strict zero-cloud architecture. All image reading, frame decoding, SAM 2.1 mask generation, optical flow tracking, and PyTorch training run 100% locally on your machine. Your images, annotation coordinates, and trained model weights never leave your local hard drive. The only network call SentinelLabeler ever makes is an encrypted license check to verify subscription status.
You own 100% of everything you produce. Unlike cloud platforms that reserve rights to train foundation models on your uploaded images or hold your datasets hostage behind paywalls, all datasets and trained PyTorch / ONNX weights are written directly to your local file system in open industry formats (Ultralytics TXT, Pascal VOC, COCO). They remain yours forever, even if you cancel your subscription.
SentinelLabeler supports all video capture hardware compatible with Microsoft Media Foundation and DirectShow. Verified devices include:
Yes, absolutely! You can drag and drop any local video file (MP4, MKV, MOV, AVI) or existing folders of static images directly into SentinelLabeler. The software handles frame extraction, deduplication, and sequential playback automatically.
Minimum Requirements (Manual Labeling):
Recommended Requirements (SAM 2 & Local CUDA Training):
SentinelLabeler executes Meta's Segment Anything Model 2.1 locally using accelerated FP16 tensor cores. When you click an object in the viewport, the SAM 2 encoder-decoder computes a sub-pixel polygon boundary around the subject in ~12ms. You can add positive points (snap more of the subject) or hold Shift to add negative points (cut out weapon models, UI overlays, or background terrain). Pressing Enter instantly converts the polygon into a clean YOLO bounding box.
In traditional tools, labeling a 5-second 60FPS clip requires drawing 300 individual bounding boxes. With SentinelLabeler's Temporal Propagation, you label the target on frame 1, press P, and our hybrid optical flow and feature-embedding tracker automatically propagates the box forward across the entire sequence. If the target ducks behind cover, the tracker pauses confidence and re-locks when they re-emerge.
SAHI (Slicing Aided Hyper Inference) breaks down ultra-wide 1440p or 4K frames into overlapping patches (such as 640×640 windows). Standard YOLO inference downscales high-res frames, causing distant character heads to turn into a handful of blurry pixels that get missed. SAHI detects micro-targets at their native pixel resolution and merges bounding boxes seamlessly across seams.
Dual-Model Live Compare allows you to load two different ONNX or PyTorch weights simultaneously. SentinelLabeler feeds the identical capture card frame to both models, rendering their detection boxes in distinct colors alongside a real-time HUD displaying confidence, IoU jitter, missed frames, and latency. This allows you to visually verify whether a newly trained checkpoint outperforms your production model before deploying it.
SentinelLabeler natively supports all modern Ultralytics and open-source YOLO architectures:
All training includes automated train/validation splitting, hyperparameter presets, live loss telemetry, and automated FP16 ONNX compilation.
No! SentinelLabeler handles everything automatically. The desktop installer provisions an isolated Python runtime and all required PyTorch CUDA 12 binaries inside %LOCALAPPDATA%\SentinelLabeler\python. You only need standard NVIDIA Game Ready or Studio graphics drivers installed.
With a single click! When local training completes, click the "Deploy to Sentinel Runtime" button. SentinelLabeler automatically exports the model to .onnx, generates the class index mapping, and copies the weights directly into the active Sentinel Core Vision models directory for immediate 1,000Hz execution.
Your license key permits activation on your personal workstations. If you upgrade your PC, build a new rig, or switch from a laptop to a desktop, you can release the old activation directly from your Account Dashboard with zero friction.
You can cancel your subscription at any time with 1 click from your Account Portal. There are no cancellation fees, no phone calls, and no awkward retention traps. Your access will continue through the end of your current billing period.
If you experience technical issues that prevent SentinelLabeler from running on your hardware and our support engineering squad cannot resolve them within 7 days of purchase, we will promptly refund your payment. Read our complete Refund Policy for full details.
Because SentinelLabeler is a newly released independent executable that is not yet signed with a multi-thousand-dollar Extended Validation (EV) certificate, Windows SmartScreen warns on first launch. The application is 100% clean and virus-free. Simply click "More info" followed by "Run anyway" to install.
Join our active community of computer vision researchers and developers on Discord.