Learn
Jev Box: Locate an Image Target
Box is an experimental Jev 1.14 question returning a rectangle. Check presence separately: an absent target still receives a box.
Quick answer
Box is an experimental Jev 1.14 question returning a rectangle. Check presence separately: an absent target still receives a box.
Request a location
Use jev-1.14.0 or jev-preview, an image, and TypeSafe-Danger-Zone-Images: true. Add these questions to the image request:
{
"parcel_present": {
"type": "noul",
"instructions": "Is a delivery parcel visible in `scene`?"
},
"parcel_location": {
"type": "box",
"instructions": "The delivery parcel nearest the front door in `scene`"
}
}There is no criteria field. Use a specific target description when multiple objects might match. Ignore the location when the presence check fails your tested policy.
Read coordinates
The answer contains box.left, box.top, box.right, box.bottom, and a separate confidence. Horizontal edges scale by image width; vertical edges scale by height. Validate ordering and bounds before cropping. One question returns one rectangle.
In a 1000 by 600 image, left 0.1 and right 0.4 correspond to x=100 and x=400. This is arithmetic, not measured model output. Confidence cutoffs require evaluation on your own images.
Evaluate before automating
Include empty scenes, multiple matches, occlusion, and small targets. Measure presence false positives and rectangle overlap against human labels. Preserve source-image dimensions when displaying a resized preview. Verify outcomes independently before using a location for an automated action.
This guide documents an experimental contract, not a live benchmark or production-readiness claim.