diff --git a/cv/Dockerfile.base b/cv/Dockerfile.base
index 1b1d184..c01cf08 100644
--- a/cv/Dockerfile.base
+++ b/cv/Dockerfile.base
@@ -62,6 +62,17 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     && pip install -r /tmp/req.txt \
     && rm /tmp/req.txt
 
+# The nvcr.io base ships its own OpenCV (pip package name `opencv`, built
+# WITHOUT FFmpeg) which shadows the opencv-python wheel from req.txt — the
+# imported cv2 ends up FFMPEG: NO, so cv2.VideoCapture can't open RTSP and
+# every camera open fails instantly. Drop the base build and force-reinstall
+# the pip wheel (FFMPEG: YES), then assert the backend is present so this
+# regression cannot return silently in a future base bump.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip uninstall -y opencv opencv-python opencv-python-headless || true \
+    && pip install --force-reinstall --no-deps "opencv-python>=4.10" \
+    && python3 -c "import cv2; bi=cv2.getBuildInformation(); assert any('FFMPEG' in l and l.strip().endswith('YES') for l in bi.splitlines()), 'OpenCV FFMPEG backend missing — RTSP capture would fail'; print('cv2', cv2.__version__, 'FFMPEG: YES')"
+
 # Diagnostic: print the compiled arch list so a future regression is
 # visible in build logs. With the nvcr.io base on 25.04, expect sm_70
 # through sm_120 (Volta through Blackwell). Pascal (sm_61) is not