OpenCV

Each warp resamples the previous warp's output, so interpolation blur compounds and you pay three passes over the image. Compose the matrices into a single 3x3 and warp once, or precompute cv2.remap maps when the transform is fixed.

Use LAB when the task is about perceived difference or about changing brightness without shifting colour: its L axis is lightness and a/b are colour-opponent axes, and distances approximate perceived difference. HSV is better for carving out a band of the colour wheel with cheap thresholds.

Processing is slower than the camera's frame rate, so decoded frames queue up behind the loop and VideoCapture.read() keeps handing back the oldest one. Latency grows without bound. Fix it by decoupling capture from processing — read in a thread that keeps only the newest frame, or drop frames with grab().

Choose the DNN module when OpenCV is already in the build, the model is a small vision network, and one dependency matters more than peak throughput. Choose a dedicated runtime when you need broad operator coverage, vendor accelerators, or models it cannot import.

Trade matching robustness against compute and memory. SIFT gives the best repeatability under scale and viewpoint change but is slowest with 512-byte float descriptors; ORB is fastest with 32-byte binary descriptors; AKAZE sits between. Decide on measured inlier ratio over your own imagery, not reputation.

Run the expensive detector every N frames and carry boxes between detections with cheap trackers, sizing N so the average per-frame cost fits the frame interval. The cost of a larger N is tracker drift and delayed discovery of new objects, so N is measured, not guessed.