v4.6.0

v4.6.0 is a major feature release. Headlining it: the camera module was renamed from sensor to csi (with a backwards-compatible sensor alias), a new ml.postprocessing module, support for the Prophesee GenX320 event sensor, PAG7936, and FLIR Boson thermal cameras, a new TinyUSB-based DFU bootloader, Nema GPU acceleration, and MicroPython 1.24. It also changes the drawing alpha range, so read the breaking changes below.

Highlights

  • sensor is now also csi — the camera module was renamed to csi; sensor remains as a backwards-compatible alias (no code change required yet).

  • New ml.postprocessing module — fomo_postprocess, yolo_v2_postprocess, yolo_v5_postprocess.

  • New sensors — Prophesee GenX320 event camera, PAG7936, and FLIR Boson 320/640 thermal.

  • New DFU bootloader — a TinyUSB-based standard-DFU bootloader replaces the legacy CDC bootloader on all boards.

  • Nema GPU — accelerated 2D/vector graphics on STM32.

  • MicroPython updated to 1.24.0.

  • Breaking: the drawing alpha range changed from 0–256 to 0–255 (see the alpha-range change).

New features

  • ml.postprocessing — a new module decoding model output: a FOMO decoder (fomo_postprocess, find_blobs + NMS, configurable threshold) and yolo_v2_postprocess / yolo_v5_postprocess (with threshold, anchors, nms_threshold, nms_sigma).

  • Event-camera visualization — new Image.to_evt_dark() / Image.to_evt_light() methods and PALETTE_EVT_DARK / PALETTE_EVT_LIGHT color palettes.

  • Depth visualization — new Image.to_depth() method and a PALETTE_DEPTH color palette (ToF-enabled boards).

  • Sensor introspection — new IOCTL_GET_RGB_STATS ioctl returning an (R, Gb, Gr, B) channel-statistics tuple.

  • GenX320 bias presets — new IOCTL_GENX320_SET_BIASES ioctl with GENX320_BIASES_DEFAULT / LOW_LIGHT / ACTIVE_MARKER / LOW_NOISE / HIGH_SPEED presets.

  • omv.debug_mode() — a new function to query whether the USB debug interface is active.

  • Nema GPU — Cortex-M55 NemaGFX/NemaVG driver integration on STM32, with in-place GPU cut-through for draw_image.

  • PAG7936 — added color-bar output, auto-exposure / auto-gain, sleep, and MIPI mode.

Other changes and improvements

  • MicroPython updated to 1.24.0; ulab updated to 6.7.3 (adds keepdims).

  • Common boot script — STM32 / i.MX RT / RP2 now share a _boot.py that auto-creates the flash/SD filesystem with a default main.py and README.txt and frees temporary startup globals.

  • STM32N6 groundwork — basic STM32N6 support, an XSPI driver, a common MPU driver, and firmware/bootloader signing on STM32.

  • Computed-goto VM — enabled on STM32 and nRF for a interpreter speedup.

Bug fixes

Camera and sensors:

  • Fixed the OV5640 PCLK/exposure calculation (two MIPI lanes, corrected divider table).

  • The MT9V0xx Bayer-CFA variant now accepts RGB565/BAYER/GRAYSCALE and outputs proper raw Bayer data instead of erroring.

  • Enabled PAG7936 lens-shading correction and fixed HD image shifting; set_framerate() is now capped per resolution.

  • Fixed FLIR Lepton thermal capture on the i.MX RT1060 (corrected VoSPI CPOL/CPHA), and Lepton snapshots now return specific error codes.

  • Fixed GenX320 frequency rounding.

Image processing:

  • Fixed memory corruption in to_ndarray() on grayscale images.

  • Fixed debayer bounds prediction in Bayer→RGB/grayscale conversion.

  • Fixed draw_image() transpose on boards without SDRAM (chunked buffer).

  • Fixed boolean-argument handling in Image.binary() and the ml.Model load_to_fb flag.

System:

  • Fixed SPI RX-only / TX-only pin configuration on STM32 and i.MX RT.

  • Moved the Portenta H7 Ethernet DMA buffers into SRAM3 so Ethernet works.

  • Fixed a dropped character in debug-channel text output over USB; rounded ML dynamic-allocation sizes up to avoid undersized model buffers.

Hardware and board support

  • Prophesee GenX320 event-based vision sensor (STM32F7 + i.MX RT1060).

  • PAG7936 sensor (with MIPI mode).

  • FLIR Boson 320/640 thermal cameras (IDD 4 driver).

  • Nema GPU acceleration on STM32.

  • STM32N6 bootloader/XSPI groundwork.

  • VL53L5CX — added power-enable-pin support.

Breaking API changes

User-visible API breaks between v4.5.9 and v4.6.0. Scope: Python C-modules in modules/ and Python libraries in scripts/libraries/.

Each change is tagged with its impact:

  • major — most scripts that used it need edits.

  • minor — narrow API; only affects scripts that used it.

  • behavior — same API, different results; re-check tuned scripts.

  • tooling — only affects flashing / host tools / board availability.

Changes are grouped by impact in that order. If you just want to port your code, jump to the migration checklist at the end. Each commit hash links to its diff on GitHub. Note: the sensor → csi module rename in this release is not breaking — sensor still works as an alias.

Drawing alpha range is now 0–255 (major)

The alpha argument of draw_image() and every overlay that takes one (FIR, ToF, display, TV, MJPEG, alpha tables) now uses a 0–255 range instead of 0–256. Scripts that passed 256 for fully opaque, or computed alpha against 256, must be rescaled to 0–255.

Commit: 1f87b5bb6

FLIR Lepton ioctl constants renamed (minor)

Three Lepton ioctl constants were shortened: IOCTL_LEPTON_SET_MEASUREMENT_MODE → IOCTL_LEPTON_SET_MODE, IOCTL_LEPTON_SET_MEASUREMENT_RANGE → IOCTL_LEPTON_SET_RANGE, IOCTL_LEPTON_GET_FPA_TEMPERATURE → IOCTL_LEPTON_GET_FPA_TEMP. The bundled examples were updated.

Commit: 12582d54d

ML post-processing API churn (minor)

ml.postprocessing is new in this release and its API moved during it: the yolo_v2_postprocess constructor keyword score_threshold was renamed to threshold, and empty post-processing results now return () instead of []. The bundled model yolov5_rgb_person was renamed to yolo_v5_224_nano. ml.apps.MicroSpeech now forwards arbitrary **kwargs to audio.init() instead of a fixed gain_db=24.

Commits: d67bd1ad7, 979b4e8ee, b89b815b2, 6671d9df5, 5c2f4388c

AGAST is the default corner detector (behavior)

find_keypoints() now uses AGAST instead of FAST as its default corner detector on all boards. Keypoint sets (and therefore descriptor matches) differ from previous releases — re-check tuned keypoint pipelines. IDE frame-buffer streaming is also now capped at 20 Hz to reduce USB load.

Commits: da6bf6910, bc4e39246

Bootloader, removed boards and features (tooling)

The legacy CDC bootloader was replaced by a new TinyUSB-based standard DFU bootloader on all boards — the firmware-update mechanism changed; update your flashing workflow/tooling accordingly. The OpenMV Pro board, the standalone Raspberry Pi Pico build, and the WiFi-debug (wifidbg) remote-IDE feature were removed.

Commits: a03fc90d8, c2e616aae, af3ea774b, f8bca9799, 59a38a0d3, 7edeb4fbe, d58d77fa2

Migration checklist

For a clean port to v4.6.0 the typical work is:

  1. Rescale every drawing/overlay alpha value from the 0–256 range to 0–255 (the alpha-range change).

  2. Rename any FLIR Lepton ioctl constants to their shortened forms (the Lepton ioctl rename).

  3. If you used the new ML post-processors, switch score_threshold to threshold, handle a () empty result, and update the yolo_v5_224_nano model name (the ML changes).

  4. Re-check find_keypoints()-based pipelines for the AGAST detector change (the corner-detector change).

  5. Update your firmware-flashing workflow for the new DFU bootloader (the bootloader/board changes).