1. Quickstart & Architecture
SentinelLabeler is built as a zero-cloud, high-throughput desktop application engineered specifically for video game footage, high-refresh HDMI capture cards, and precision object tracking.
%LOCALAPPDATA%\SentinelLabeler\python on first launch. It never modifies your system PATH, never conflicts with existing Anaconda/Miniconda setups, and automatically configures CUDA 12.x drivers.
Installation Checklist
- Download the official Windows installer from the Downloads Page (
SentinelLabelerSetup.exe). - When prompted by Windows SmartScreen, click "More info" followed by "Run anyway" (executable is self-extracting).
- Launch SentinelLabeler. On first run, it will verify your NVIDIA GPU drivers and initialize the local PyTorch CUDA runtime (~90 seconds).
- Enter your license key obtained from your Account Portal.
2. Capture Card Setup & Ingest
Unlike generic photo annotators, SentinelLabeler communicates directly with video capture hardware via Microsoft Media Foundation and DirectShow for ultra-low latency frame acquisition.
Supported Hardware
- Elgato HD60 X / 4K X / 4K60 Pro: Supports up to 240Hz input at 1080p and 120Hz at 1440p.
- Magewell USB Capture HDMI Gen 2: Zero-driver plug-and-play DirectShow streaming.
- AverMedia Live Gamer ULTRA 2.1 / 4K: High-bandwidth YUY2 & NV12 pixel ingestion.
- Local Video Files: MP4, MKV, MOV, and AVI containers up to 4K resolution.
3. SAM 2.1 Auto-Masking & Text Prompts
SentinelLabeler integrates Meta's Segment Anything Model 2.1 (SAM 2.1) alongside YOLO-World for lightning-fast zero-shot bounding-box and segmentation mask generation.
Interactive Click-to-Mask
Simply select the SAM 2 tool or press S, then left-click directly on any character, vehicle, or object in the viewport. SAM 2.1 will compute the exact boundary polygon in under 15 milliseconds.
- Positive Point: Left Click adds an inclusion point to snap the mask to the target.
- Negative Point: Shift + Left Click places an exclusion point to subtract background clutter, shadows, or weapon silhouettes.
- Convert to Bounding Box: Press Enter to convert the SAM 2 mask into a tight, rotation-compensated YOLO bounding box.
Open-Vocabulary Text Prompting
Using YOLO-World zero-shot prompting, enter colloquial terms into the prompt bar (such as enemy soldier helmet, vehicle wheel, or crosshair center). SentinelLabeler will scan incoming video frames and automatically generate candidate labels matching your descriptions without any pre-training.
4. Temporal Video Flow & Frame Propagation
Manual frame-by-frame labeling is exhausting. With SentinelLabeler's Temporal Video Flow engine, you label a single keyframe and let the temporal tracker carry your bounding boxes forward across hundreds of consecutive frames.
5. SAHI High-Res Tiled Inference
Standard YOLO models typically downscale inputs to 640×640 pixels, destroying the details of distant characters or tiny heads across a 1440p or 4K capture feed. SentinelLabeler solves this with SAHI (Slicing Aided Hyper Inference).
How SAHI Works
SAHI slices the full-resolution frame into overlapping sub-tiles (e.g., 640×640 patches with a 20% overlap ratio), runs batch inference on all slices simultaneously using TensorRT / ONNX Runtime, and merges duplicate detections using Non-Maximum Suppression (NMS). This allows detection of micro-targets as small as 8×8 pixels across ultra-wide canvases.
6. Dual-Model Live A/B Compare
How do you know if your newly trained yolo11n_custom.pt is actually better than your baseline yolov8s_production.onnx? SentinelLabeler includes a real-time side-by-side comparative inspection HUD.
Comparative Telemetry
- Side-by-Side Viewport: Model A rendered on the left feed, Model B rendered on the right feed with synchronized timeline playback.
- Box IoU Jitter Metric: Measures frame-to-frame bounding box shudder; lower jitter means smoother tracking for downstream visual servoing.
- Confidence Delta: Real-time histogram showing which model identifies occluded targets faster with fewer false positives.
- Inference Latency: Millisecond benchmark timer measuring exact GPU execution cost per frame.
7. Local CUDA Model Training Studio
No need to export datasets to Google Colab, pay for cloud GPU hours, or configure complex Linux virtual environments. SentinelLabeler contains an integrated PyTorch training harness built for Ultralytics YOLO models.
Supported Architectures
SentinelLabeler natively supports training on:
- YOLO11: Nano (
yolo11n), Small (yolo11s), Medium (yolo11m) — optimal latency/mAP balance. - YOLOv10 & YOLOv9: NMS-free dual-label assignment models.
- YOLOv8: Battle-tested production industry standard.
Training Configuration (data.yaml)
During training, the Studio renders live interactive loss curves (Box Loss, Class Loss, DFL Loss) and validation mAP50 metrics. When training finishes, models are automatically exported to both best.pt and optimized FP16 best.onnx.
8. Dataset & Weight Export
Your data belongs to you. SentinelLabeler provides instantaneous zero-lock-in exports to all major machine learning formats:
- Ultralytics YOLO (TXT): Normalized coordinates
<class-id> <x_center> <y_center> <width> <height>. - Pascal VOC (XML): Pixel bounding box coordinates
<xmin> <ymin> <xmax> <ymax>. - COCO JSON: Polygon segmentation annotations compatible with Mask R-CNN, Detectron2, and MMDetection.
- Sentinel Core Vision Hot-Swap: Direct 1-click deployment into the Sentinel 1,000Hz runtime folder.
9. Keyboard Shortcuts & Hotkeys
Master these high-speed keyboard shortcuts to label footage at over 1,000 frames per hour:
| Hotkey | Action | Description |
|---|---|---|
| Space | Auto-Suggest Box | Runs active YOLO model inference on current frame to place candidate boxes. |
| W | Create Bounding Box | Activates crosshair cursor to drag-draw a new bounding box. |
| S | SAM 2 Masking | Toggles Segment Anything 2.1 click-to-mask mode. |
| P | Temporal Propagate | Propagates active boxes forward through subsequent video frames. |
| A / D | Prev / Next Frame | Steps timeline by exactly 1 video frame. |
| Shift + D | Jump +10 Frames | Skips forward 10 frames for rapid footage scrubbing. |
| 1 – 9 | Class Selector | Instantly sets the active target class ID for subsequent boxes. |
| Delete | Remove Box | Deletes the currently selected bounding box or mask. |
| Ctrl + Z | Undo | Reverts last annotation or vertex modification. |
| Ctrl + S | Save Labels | Immediately commits label TXT file to disk. |
10. Troubleshooting & System Logs
If you encounter any issues during capture card streaming or model training, consult the following diagnostic checklist:
1. Windows SmartScreen Warning on Install
Because SentinelLabeler installers are self-extracting and not yet EV code-signed, Windows Defender may display a blue banner reading "Windows protected your PC". Simply click "More info" and select "Run anyway".
2. Black Screen on HDMI Capture Card
Ensure that HDCP (High-bandwidth Digital Content Protection) is disabled on your video source (PlayStation 5, Xbox Series X, or Nintendo Switch). If using an Elgato card, verify that the 4K Capture Utility is closed before launching SentinelLabeler, as Windows allows only one application to lock the DirectShow video feed at a time.
3. CUDA Out of Memory (OOM) During Training
If local training halts with a PyTorch CUDA Out-Of-Memory error, reduce the Batch Size from 16 to 8 or 4 in the Training Studio settings, or reduce the image input resolution from 1280 to 640.
4. Diagnostic Log Files
SentinelLabeler writes detailed timestamped logs to:
If you need assistance, submit this file to our technical support engineers on Discord or via [email protected].