Object Detection
EfficientDet-Lite0/2 with per-class confidence scores, NMS and bounding-box tracking across frames.
- 80 COCO classes
- Custom
.tfliteupload - Class filters
Benzo Vision turns your webcam into a real-time perception engine — object detection, pose estimation, hand tracking, gesture recognition and face landmarks. Every frame is processed locally. Nothing ever leaves your device.
Each task shares the same capture pipeline, GPU delegate and overlay renderer. Switch models without reloading the page.
EfficientDet-Lite0/2 with per-class confidence scores, NMS and bounding-box tracking across frames.
.tflite upload33 3D landmarks with visibility scores. Compute joint angles, rep counts and form feedback in real time.
21 landmarks per hand, up to two hands, with left/right handedness. Drives the touchless cursor.
478 points including irises, plus 52 ARKit-style blendshapes for expression-driven effects.
Drop in a .tflite or point at a hosted model URL. It's hashed, cached in IndexedDB and reused offline.
Record the annotated canvas to WebM or grab PNG snapshots — overlays are rendered into the pixels.
The capture loop never blocks the UI thread. Inference runs in a module worker; overlays are drawn on the main thread against a transferred result object.
getUserMedia / getDisplayMedia / file input feeds a hidden <video>.
createImageBitmap() snapshots the frame and hands it to the worker as a transferable.
MediaPipe Tasks runs on the GPU delegate (WebGL / WebGPU) with WASM SIMD fallback.
Landmarks and boxes are projected and stroked onto the overlay canvas at device resolution.
Gesture events become synthetic clicks and swipes for touchless UI control.
Model blobs land in IndexedDB so the second load is instant and fully offline.
All models are cached after first download. Sizes shown are quantized weights.
| Model | Task | Size | Latency* | License |
|---|---|---|---|---|
| EfficientDet-Lite0 | Object detection | 4.4 MB | 9 ms | Apache-2.0 |
| EfficientDet-Lite2 | Object detection | 7.2 MB | 22 ms | Apache-2.0 |
| Pose Landmarker Lite | Pose | 5.8 MB | 11 ms | Apache-2.0 |
| Pose Landmarker Full | Pose | 12.1 MB | 26 ms | Apache-2.0 |
| Hand Landmarker | Hands | 7.8 MB | 13 ms | Apache-2.0 |
| Face Landmarker | Face + blendshapes | 3.8 MB | 8 ms | Apache-2.0 |
| Your model | Any | — | — | Yours |
*Measured on an M1 MacBook Air, 640×480 input, GPU delegate, median of 300 frames.
No sign-up, no server, no telemetry. Open the studio and start detecting in seconds.