JavaCV for Vision on the JVM

JavaCV Is the Camera + OpenCV Default

• For camera capture, video I/O, and OpenCV operations on JVM, use JavaCV (org.bytedeco:javacv-platform, current 1.5.13)
• JavaCV bundles JavaCPP wrappers for OpenCV, FFmpeg, libdc1394, OpenKinect, ARToolKitPlus — a single dependency for the whole stack
• Don't shell out to cv2 via Python, don't try to use org.openpnp:opencv (older, abandoned), don't write JNI bindings to a system OpenCV install

What You Get

• org.bytedeco.javacv.OpenCVFrameGrabber(index) for webcam capture
• org.bytedeco.opencv.global.opencv_imgproc.* for full OpenCV imgproc API (cvtColor, resize, rectangle, putText)
• org.bytedeco.opencv.opencv_objdetect.CascadeClassifier for Haar detection
• org.bytedeco.javacv.Java2DFrameConverter + OpenCVFrameConverter.ToMat for Mat ↔ BufferedImage ↔ DJL Image bridging

Idiomatic Use

• Wrap OpenCVFrameGrabber in an AutoCloseable Kotlin class with try { ... } finally { grabber.stop() }
• On macOS, the first grabber.grab() after start() returns black frames for ~500 ms — sleep through the warm-up before entering the main loop
• Combine with a Flow<Frame>: emit on every grab, use Flow.sample() to throttle heavy inference downstream

Footprint Warning

• javacv-platform pulls native libs for Linux x86_64, Linux ARM64, macOS x86_64, macOS ARM64, Windows x86_64 — ~400 MB of jars on first download
• Pre-warm Gradle dependency cache before any environment with constrained bandwidth
• For production deployments to a single platform, use the platform-specific artifact (javacv + opencv-platform-linux-x86_64) instead of the full javacv-platform

Anti-patterns

• ❌ Shelling to python -c "import cv2; …" for image work — defeats the point of being on JVM
• ❌ org.openpnp:opencv for new code — unmaintained, lacks recent OpenCV features
• ❌ System libopencv.so via JNI hand-bindings — JavaCV solves this; don't re-solve it
• ❌ Re-allocating Mat per frame inside a 30 fps loop — reuse the buffer