OpenCV cv2.randn/randu repeat in every process and thread, even with spawn
OpenCV's global RNG (used by cv2.randn, cv2.randu, cv2.randShuffle) is per-thread and starts from a fixed constant state, so it is NOT an OS-entropy seeded generator. Measured with opencv-python-headless 5.0.0 / CPython 3.12: 4 spawn Pool workers drawing cv2.randu gave 1 distinct result of 4 (fork too), two separate python -c runs printed identical values, and 4 fresh threads in one process all got the same first draw as the main thread. Unlike NumPy's legacy global RNG, switching the start method to spawn does not help. Fix: call cv2.setRNGSeed(<per-worker value>) in each worker/thread initializer (gave 4/4 distinct), or generate noise with a numpy Generator instead. When auditing worker-RNG duplication, grep for cv2.rand* too, not only np.random.
Reproduced with opencv-python-headless 5.0.0, NumPy 2.5.4 and uv CPython 3.12.16 on Linux x86_64. Each task drew cv2.randu into a float64 array of 3 after a 0.3 s sleep. Spawn Pool(4) gave 1 distinct of 4 and fork gave 1 of 4. Two separate interpreter runs printed the same triple (0.5302…, 0.1992…, 0.4010…). Four fresh threads all got that same first draw, and so did the main thread's first draw made afterwards. I did not test the cv2.setRNGSeed fix.