Segmentation mask
Per-frame binary mask of the animal body at pixel resolution. Enables custom downstream analysis of body shape features.
PNG seriesNPY
← Capabilities
Segmentation-Based Detection
Pixel-precise animal detection through deep image segmentation
U-Net segmentation produces exact body contours regardless of arena contrast — no threshold tuning, no background calibration, no lost animals on dark bedding.
98%
Segmentation IoU accuracy
30fps
Real-time inference speed
0
Manual threshold tuning needed
1px
Body boundary precision
The problem
Threshold-based detection fails when contrast is low
Traditional tracking relies on brightness thresholds to separate the animal from the background. When a dark animal walks on dark bedding, or a white animal is tested on a light surface, threshold detection fails — producing tracking dropouts, fragmented detections, and lost position data.
- Dark-on-dark and light-on-light scenarios cause complete tracking loss
- Bounding-box detectors discard body shape information needed for posture and orientation analysis
- Manual threshold adjustment wastes time and varies across sessions and scorers
The solution
Learned segmentation that works at any contrast
ConductVision uses a U-Net segmentation model trained on diverse animal/arena combinations. The model produces pixel-level animal masks regardless of contrast, providing exact body contour, area, elongation, and orientation.
- Works on any contrast combination — no threshold tuning required
- Pixel-level body contour enables area-based freezing detection and posture analysis
- Body elongation and orientation computed directly from segmentation mask geometry
Endpoints
Segmentation-derived measures
Body morphometrics
Body area, perimeter, elongation ratio, and major/minor axis lengths per frame — derived from mask geometry.
CSVJSON
Contrast-independent position
Centroid coordinates from segmentation mask center-of-mass — robust position tracking even in zero-contrast conditions.
CSV
Applications
Applications requiring robust detection
Area-based freezing quantification
Body area change between frames detects freezing more sensitively than centroid displacement — the animal can sway without locomoting.
Measures
- Body area change rate
- Freezing bout detection
- Motion vs. area agreement
Dark bedding tracking
Black mice on dark corn cob bedding — a common scenario that breaks threshold tracking. Segmentation maintains detection.
Measures
- Detection rate
- Tracking continuity
- Position accuracy
Posture classification from contour
Body elongation ratio distinguishes stretched-attend posture from ball posture — an ethologically relevant anxiety indicator.
Measures
- Elongation ratio
- Posture state classification
- Stretched-attend frequency
Segmentation-based identity separation
Individual animal masks enable identity maintenance through close interactions where bounding boxes merge.
Measures
- Identity preservation rate
- Separation during contact
- Individual body metrics
Compared to typical systems
How ConductVision differs
| Feature | ConductVision | Typical systems |
|---|---|---|
| Low-contrast detection | Works at any contrast level | Requires contrast threshold tuning |
| Body shape data | Pixel-level contour | Bounding box only |
| Background calibration | Not required | Required before each session |
| Posture features | Elongation, area, orientation | Not available from detection |
| Inference speed | 30 fps real-time | Often post-hoc only |
Related capabilities
Precision Tracking
High-resolution 30 fps tracking that captures sub-second behavioral events conventional systems miss.
Freezing Detection
Dual-mode automated freezing scoring with TTL/serial apparatus integration for fear conditioning paradigms.
Behavior Classification
ML-powered frame-level behavioral classification with temporal smoothing and custom classifier training.
Never lose tracking to poor contrast again
Upload a challenging video and see segmentation-based detection maintain tracking where thresholds fail.
