JeVois 1.21 brings the following new features:

Major upgrade to Ubuntu 24.04 (noble) for development and also running in JeVois-Pro. JeVois-A33 maintains its custom embedded operating system based on buildroot. This sets us at the forefront of the newest long-term-support (LTS) Ubuntu version, which means more years of updated packages ahead (official support until April, 2029).
Updates to OpenCV 4.10.0, Python 3.12 (on JeVois-Pro only), Numpy 1.26, Mediapipe 0.10.14 custom patched and compiled for noble/python 3.12, onnxruntime 1.18.0, Hailo 1.17, NPU ovxlib 1.37.0, updated and patched and recompiled pycoral and tflite-runtime for Python 3.12, and more. JeVois-Pro now runs Ubuntu 24.04 aarch64 with many available latest packages.
JeVois-Pro: Ubuntu noble does not like to allow users to install python packages with pip globally as they may conflict with packages installed via apt-get. Hence, now running jevoispro-daemon in a virtual env in /root/jvenv/
New docker images and scripts to recompile the whole operating system and/or machine vision modules (JeVois-A33 and JeVois-Pro), or emulate the JeVois-Pro aarch64 O.S. Note that these images run Ubuntu 24.04 and there may be compatibility issues with older hosts.
New module CalibrateCamera for camera calibration using a variety of calibration boards (chessboard, ChArUco, etc). Reworked camera calibration file naming as calibration-sensor-WxH-lens.yaml, new helper function to load/save calibration data transparently for a given video mapping, without having to compute the file name of reach the file manually.
JeVois-Pro: New serial monitors allow one to see serial communications that are not directed to the console (e.g., monitor traffic between a micro-controller and JeVois over the 4-pin hardware serial port). Very useful for debugging micro-controller code.
JeVois-Pro: Updated to using g++-14 with support for C++-26.
JeVois-Pro: New PyLLM module allows you to run mobile-friendly large language models (LLMs) in a simple Python module. Okay performance with tinydolphin (1.1 Billion parameters). Support for vision language models (VLMs) also added, though more research is needed to reduce their size and CPU demands (moondream with 1.7 Billion parameters is pre-installed but too slow to be seriously usable at this stage). New ChatBox widget to facilitate user interaction with LLMs, written in C++ with python bindings.
JeVois-Pro: New post-processors and hardware-accelerated networks for YOLOv8-segmentation, YOLOv10, etc.
JeVois-Pro: New ability to peek into neural network output data, which is useful to debug neural nets converted to various hardware accelerators. Helped us during the conversion of YOLOv10 for NPU.
JeVois-Pro: Now also creating a script with each new module to recompile it from the command-line. Useful if a code change gives rise to a hard crash (e.g., segmentation fault) which would previously make it impossible to edit the code in the JeVois-Pro GUI.
JeVois-Pro: New AprilTag module written in Python to detect April Tags using the apriltag library, as a possible alternative to using the OpenCV ArUco module (which also supports several AprilTag dictionaries).
DNN: removed support for pre/net/post types Custom; instead, provided an example CustomDNN of how to create derived classes to support customized DNN processing in C++. CustomDNN implements a derived Post-Processor class, which extracts a high-resolution region of interest (ROI) from each detection box when running YOLO or similar; it then processes each ROI (for tutorial purposes, just extracting edges) and displays the results super-imposed on top of the original ROIs.
Reworked serial inputs to avoid possible buffer overflow in the Linux kernel, which could eventually corrupt a running NPU deep net.
JeVois-A33: Updated a few buildroot packages for compatibility.
JeVois-A33: Simplified the module installation process as it was quite confusing with the staging, microSD, or live targets. We now always compile and install into /var/lib/jevois-microsd (for later flashing using jevois-flash-card) and to .jvpkg packages (which can be copied to microSD and will be unpacked by the camera on next reboot).
Fixed a bug in GPUimage which would lead to corrupted display if an image changed size from one video frame to the next.
JeVois-Pro: Rework of drawPoly() for better drawing of polygons in the JeVois-Pro GUI.
Updated handling of Python exceptions for Python 3.12.
Added shorthand version of itext() for Python, leading to much lighter syntax to display information messages from python modules.
New script in jevoispro-sdk to install non-deb dependencies for host, such as pip packages, ollama, etc. Those are all pre-installed on platform, so this script is to help development on host.
Simplified logic for jevois::Parameter::freeze(), now only one function with a bool arg.
New scripts to download and install the jevois and jevoisbase third-party contrib packages directly in pre-compiled binary form, which may eliminate possible issues while trying to recompile them from source (especially tensorflow using bazel, which tends to have lots of possible compilation problems).
JeVois-Pro: Changed all GPU shaders to use highp precision, thereby eliminating all aliasing issues previously observed with some image resolutions.
JeVois-A33: jevois-flash-card can now create a disk image, as opposed to only flashing to physical microSD card.
Miscellaneous bug fixes and performance improvements.
Many doc updates.