Computer Vision

How an image becomes an array, why OpenCV hands you BGR, and what uint8 does at 256.

imread returning None instead of raising, headless imshow, and checking every frame you are handed.

Blur, gradients, adaptive thresholding, and the lighting change that makes all three collapse.

Cleaning a binary mask, then reading area, centroid, and orientation off the shapes you kept.

Composing the steps into one tool with tunable parameters and a visual dump of every stage.

Projection, focal length, exposure, and why your lens bends every straight line near the edge.

Print a board, solve for intrinsics and distortion, and reject a calibration by its reprojection error.

ORB keypoints, matching, RANSAC, and flattening a surface you photographed at an angle.

Scale from a known reference, pose from fiducial markers, and quoting your error instead of hiding it.

Fine-tune a pretrained backbone on a few hundred images; you will not train from scratch.

Grounding DINO and SAM 3 finding objects nobody labeled, then the prompts where they quietly fail.

Boxes against masks, anchors against queries, NMS-free heads, and reading mAP without flattering yourself.

AGPL and its network clause, Apache weights, non-commercial checkpoints, and picking what you can actually deploy.

Auto-label with a foundation model, correct by hand, and audit what your camera never saw.

Near-duplicate video frames, perceptual hashing, and why your 0.95 mAP is fake.

Decoding, dropped frames, buffering, and processing at capture rate instead of falling further behind.

IDs across frames, recovery after occlusion, and counting over a line without double-counting.

PyTorch to ONNX to INT8, choosing calibration data, and measuring the accuracy you traded away.

Profile on a Pi, a Jetson, or a laptop, and discover the bottleneck is your resize, not the model.

Domain shift, subgroup error rates, faces, and the deployment you decline to build.