3.4.0Runtime
Overrun policies per node
Each node in a graph can now choose what happens when it misses its slice of the frame budget.
- NewPer-node overrun policy: reuse last result, skip, or degrade to a smaller model variant.
- NewThermal model for Ambarella CV72S, so schedules keep their slack under sustained load.
- ImprovedDepth-M on Jetson Orin NX is 9% faster after fusing the mask head into the detector.
- FixedRare double frame on camera reconnect when the image processor restarted mid-stream.
3.4.0SDK
Async streams in Swift and Kotlin
Events arrive as native async sequences and flows, with back-pressure handled by the Runtime.
- NewSwift: session.events is an AsyncSequence. Kotlin: a cold Flow.
- NewPython wheels for Linux aarch64 with Jetson and Hailo back ends.
- ImprovedC++ headers compile cleanly with -Wall -Wextra on GCC 13 and Clang 17.
2.6Studio
Trace compare
Open two traces side by side and see where each frame spent its budget.
- NewTrace compare view with per-node timing deltas and frame-by-frame scrubbing.
- NewExport a trace as a reproducible benchmark run for parallax bench.
- ImprovedZone editor snaps to floor markings detected in the first frame.
2.2Depth models
Depth-Event
A motion model that reads event-camera streams directly. Available on Apple Silicon, Hexagon and Jetson.
- NewDepth-Event: optical flow from event voxel grids, 4.1 ms on Hexagon.
- ImprovedDepth-Pose keypoint AP up 1.4 points on crowded scenes.
3.3.1Runtime
Stability release
Fixes for long-running sessions on battery-powered devices.
- FixedMemory growth of about 2 MB per hour when tracks were dropped and re-acquired.
- FixedClock drift between camera timestamps and event timestamps after sleep on Apple Silicon.
3.3.0Runtime
Exposure-banded int8
Quantization scales now follow the exposure the image processor reports. Low-light recall improves with no change in latency.
- NewExposure-banded calibration on by default for every int8 model.
- Newparallax compile --calibrate exposure-bands=N for custom models.
- ImprovedZero-copy hand-off from the image processor on Hailo-8 host boards.
2.5Studio
Two-minute camera calibration
Add a camera, record a short capture, and Studio picks the model variant matched to that sensor.
- NewSensor identification from a calibration capture.
- NewFleet view for up to 64 cameras on one site.
- FixedOverlay labels clipped at the right edge of 4:3 feeds.
3.2.0SDK
One API across four languages
Swift, Kotlin, C++ and Python now share the same session, camera, model and event types.
- NewKotlin bindings with Android camera2 capture built in.
- ImprovedPython session.watch() yields typed events instead of dictionaries.
3.2.0Runtime
Ambarella support
The Runtime compiles schedules for CVflow, bringing the full Depth family to Ambarella CV72S.
- NewAmbarella CV72S target for parallax compile.
- ImprovedGraph compile time down from minutes to under 40 seconds for typical graphs.
3.0.0Runtime
The frame-budget scheduler
Runtime 3 replaces the request queue with a per-frame budget. Every node gets a slice, and nothing waits.
- NewFrame-budget scheduler with stale-result hand-off.
- NewOne compiled schedule per chip instead of per framework.
- Improvedp99 latency on Hexagon down from 41 ms to under 10 ms on the reference graph.