Skip to content

Release v3.12.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Nov 06:23
v3.12.0
3c28296
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Features

  • Multi-stream PointCloud

    • PointCloud now merges any number of depth streams (each with optional aligned color) into one PointCloudData. Add streams with getDepthInput(name) and getColorInput(name)
    • setTargetCoordinateSystem(deviceId, ...) puts the point cloud in the coordinate system of a camera or the housing of any device in the pipeline
    • [RVC4] useGPU() moves the deprojection to the GPU to reduce the CPU load. The CPU stays the default, because it is faster
    • A Python example and a C++ example are available

  • [Beta promotion] Stitching node

    • Now available as dai.node.Stitching, also in builds without beta features
    • Calibrated panorama stitching is now the default and needs undistorted inputs. Use setUseInputCalibration(false) for feature-based stitching
    • dai.beta.node.Stitching is deprecated and will be removed in a future release
    • Multi-device panorama examples are available in Python and C++

  • [Beta promotion] ImgDetectionsFilter node

    • Now available as dai.node.ImgDetectionsFilter. Filters detections by label, confidence, size, region of interest and count, also at runtime through inputConfig
    • Merges detections from multiple cameras into one reference image (for example a stitched panorama) and removes duplicates with NMS or AVERAGE
    • Runs on RVC4 with one input. With multiple inputs, and on RVC2, it runs on the host
    • The interface changed (inputs[name] and out replace input and output). dai.beta.node.ImgDetectionsFilter is deprecated and will be removed in a future release
    • A single-camera example and a multi-device panorama example in Python and C++ are available

  • Resize mode from NN archives

    • NeuralNetwork, DetectionNetwork and SpatialDetectionNetwork built from a camera now use the resize_mode (CROP, STRETCH or LETTERBOX) declared in the NN archive. Archives without it use CROP, as before
    • Override it with the resizeMode argument of build

  • [RVC2] Multi-device calibration (pipeline.setMultiDeviceCalibration) is now supported on RVC2 devices

  • Align.setInterpolation(...) selects the interpolation (BICUBIC, BILINEAR or NEAREST_NEIGHBOR). AUTO keeps the previous behavior

  • New dai.node.Depth.getSupportedAlgorithms(device) and getDefaultAlgorithm(device) return the depth algorithms that a device supports and the one that AUTO selects

  • New Device.getProtocol() returns the connection protocol (USB or TCP/IP)

  • [RVC4] setAutoFocusTrigger() now runs one focus scan and then holds the focus. The AF mode is reset for each session

  • [RVC2] New Device.Config.dangerousIrPowerOverride option removes the combined dot and flood power limit on boards that have one. Full power on both emitters can crash the device

Bug fixes

  • Alignment
    • ImageAlign and Align now keep the depth near the image edges for rotated and distorted cameras
    • [RVC4] Align now uses correct offsets for rotated cameras and keeps the configured output size
    • ImageAlign no longer crashes when the inputAlignTo frame has no distortion coefficients
    • ImgDetections.transformTo() no longer distorts boxes at some rotations when mapping to a non-square image
    • [RVC2] StereoDepth depth alignment is now correct after decimation (for example on OAK-D-SR)
  • [RVC4] Stability
    • Resolved a firmware crash when a device is opened right after the previous session closes (USB, OS >= 1.41.0)
    • Resolved crashes on USB session close, at StereoDepth stop, at VideoEncoder stop with MJPEG decoder input, and in the logger at shutdown
    • Resolved crashes in ImageManip crops with fractional dimensions and in FeatureTracker with an invalid input
    • Fix for GPU memory corruption
    • ImageManip no longer adds a black border around a cropped rotated rectangle
    • crash dump retrieval retries when the dump is not ready yet
  • [RVC2] Stability
    • On OAK-D Pro PoE, the dot projector and the flood light now share one intensity budget of 1.0 (for example 50% dot + 50% flood). Before, both emitters at full intensity could draw too much power and crash the device
    • Resolved a rare firmware crash in the PoE network stack under heavy TCP traffic
    • Resolved warp timeouts and all-zero output frames caused by the hardware filter interrupt handling
    • Camera-module EEPROM access no longer resets the other sensor on a shared I2C bus (for example OAK-FFC-3P)
    • RVC2 crash dumps are now saved when device strings are not valid UTF-8
  • [RVC4 OS >= 1.47.0] depthai-gate v0.0.30
    • Resolved issue where firmware would crash because gate would accept new session before the old process had exited
    • USB transfers now have a time limit, and the gate recovers after a failed transfer, so the USB connection no longer stalls
    • A session that starts late is no longer reported as a crash when the device closes
    • When the firmware cannot start, the host now gets an error instead of no response
    • Old crash dumps are deleted when a new session starts
  • ObjectTracker srcImgDetection now keeps the original normalized coordinates
  • NeuralAssistedStereo with rectifyImages=false no longer rectifies the frames
  • pipeline.start() now throws an exception instead of aborting the process when the device link drops during the asset transfer
  • [PoE/TCP] The host now closes the socket after a connection reset, so many reconnects no longer cause "Too many open files"
  • Resolved possible memory corruption in XLink after a read timeout
  • A USB boot with an incomplete firmware upload now fails at once instead of waiting for the boot timeout
  • Device.performHealthCheck() now stops on Ctrl+C and accepts a shouldStop callback
  • Holistic record requires the video encoding settings only when video encoding is enabled

Performance

  • [RVC4 OS >= 1.46.0] NeuralDepth uses new optimized LENS models. Model FPS with the balanced performance profile:
    • N 384x240: 120.42 → 251.17 FPS (+109%)
    • S 480x300: 73.91 → 153.78 FPS (+108%)
    • M 576x360: 48.21 → 120.73 FPS (+150%)
    • L 768x480: 26.72 → 68.31 FPS (+156%)
    • XL 864x540: 22.00 → 52.06 FPS (+137%)
    • 2XL 960x600: 17.21 → 42.97 FPS (+150%)
    • 3XL 1056x660: 14.90 → 33.99 FPS (+128%)
    • 4XL 1248x780: 9.85 → 23.56 FPS (+139%)
  • XLink now reads up to 8 MiB messages per transfer resulting in up to 12% higher USB throughput for large messages on RVC4
  • [RVC4] Undistorted Camera outputs now use the GPU and are 3 to 4 times faster. Plain resizes are up to 4.8 times faster for large frames

Misc

  • [RVC4] The OS power profile is logged on connection, and a warning is shown when the power profile limits the IR intensity

  • Startup notifications

    • After a device connects, DepthAI shows notices that apply to your DepthAI version, platform or device, for example a new release or a bootloader update
    • Set DEPTHAI_DISABLE_STARTUP_NOTIFICATIONS=1 to turn them off

  • [RVC4] Added support for readCcmEepromRaw() and writeCcmEepromRaw()

  • NN archive loading and model zoo downloads in Python release the GIL, so other threads no longer stall while a large model loads

  • Added missing Python bindings and fixed incorrect ones

  • Added Python 3.15 wheels

[RVC4] Luxonis OS compatibility

Integration tested with Luxonis OS 1.35.0, 1.41.0, 1.45.0 and 1.46.0

Assets 9
Loading