Author

Esdras Ntuyenabo

Engineer · LexData · 31 articles

Camera on a pole over an industrial yard at dusk, camera, fence and vehicle boxed, generated scene with detections from our model

Computer vision · 6 min read

Object detection in production, what a detector returns and what it learns from

For every frame off the yard camera the detector returns a list of boxes with a class and a score, and the score is a ranking rather than a probability.

Esdras Ntuyenabo · Oct 4, 2026

Scaffold with two workers, guardrails, ladder and plank decks boxed, from a customer site camera

Computer vision · 6 min read

Transfer learning starting points, why a small nearby dataset beats a huge general one

Four checkpoints for the same hard hat model, and the small one trained on people won. Pick a base model by what it has seen, then feed it the site's frames.

Esdras Ntuyenabo · Oct 4, 2026

Transmission tower over a country road from a drone, six insulator strings and one corrosion mark boxed, from a customer inspection run

Labeling · 7 min read

Aerial dataset augmentation for drone frames where there is no up

A tower seen straight down has no top or bottom, so rotations and both flips are safe. Scale for altitude, brightness for sun, move every box with its pixels.

Esdras Ntuyenabo · Oct 3, 2026

Bread shelf with an empty slot flagged and the rack sections boxed, a shopper at the end of the aisle, from a customer store camera

Computer vision · 7 min read

Image recognition AI, explained on the cameras a site already owns

A tag, a box, a mask or keypoints, on a substation camera, a line camera and an aisle camera. How a model gets there, and what its score does not promise.

Esdras Ntuyenabo · Oct 3, 2026

Robotic welding cell with the torch and the robot boxed over the fixture, generated scene with detections from our model

Computer vision · 7 min read

Image segmentation and the questions on a weld bead that only a mask can answer

A box finds the weld. A mask measures it. Semantic, instance and panoptic told apart on a weld cell and a board camera, with what each costs to label.

Esdras Ntuyenabo · Oct 3, 2026

Transmission line over farmland with towers, insulators, defects and vegetation boxed, from a customer aerial survey

Labeling · 7 min read

COCO vs YOLO format, converted without losing a box

COCO vs YOLO format, with CVAT XML beside them. Three years of inspection labels arrive in three formats, and the boxes that shift are those nobody drew.

Esdras Ntuyenabo · Oct 2, 2026

Camera on a pole over an industrial yard at dusk, generated scene with detections from our model

Labeling · 7 min read

Image resizing for computer vision, and the resize that squashes the forklift

A wide yard frame goes into a square input. Pad rather than stretch, move the boxes with the pixels, and treat a changed interpolation on the runner as a step.

Esdras Ntuyenabo · Oct 2, 2026

Stamping line, steel sheet passing the press, generated scene with detections from our model

Operations · 7 min read

Drawing polygon regions of interest for the zone a rule applies to

The press zone, the dock lane and the checkout queue are polygons on a frame. Perspective makes them trapezoids, and a nudged camera moves them all at once.

Esdras Ntuyenabo · Oct 2, 2026

Warehouse aisle with pallets and a forklift boxed, generated scene with detections from our model

Operations · 9 min read

Convert DAV to MP4 with ffmpeg, without re-encoding

DAV is a recorder's wrapper round an ordinary H.264 stream. One ffmpeg command copies it into an MP4 untouched. The rest is for the files that refuse.

Esdras Ntuyenabo · Oct 1, 2026

Cartons and labels boxed on a packaging line, generated scene with detections from our model

Edge · 6 min read

Inference latency in computer vision, and the budget between the camera and the reject arm

A carton takes a fixed time to reach the reject arm from the lens. Everything from capture to the signal has to fit inside it, and most of it is not the model.

Esdras Ntuyenabo · Oct 1, 2026

A camera on a pole over an industrial yard at dusk, generated scene with detections from our model

Edge · 6 min read

RTSP stream processing for computer vision, and the three ways a camera feed fails without an error

A feed that falls behind, a reboot that ends the frames with no error, a stream with the wrong clock. Sample every two seconds and give each camera a pulse.

Esdras Ntuyenabo · Oct 1, 2026

A checkout lane from a ceiling camera, person and products boxed and no face in the frame, generated scene with detections from our model

Operations · 6 min read

Face blurring for privacy in computer vision, done before the frame is stored

A blur applied after the request arrives is a blur applied to a frame that has already been copied. The step belongs at ingestion, beside the recorder.

Esdras Ntuyenabo · Sep 30, 2026

Produce packhouse, apples on a grading conveyor, generated scene with detections from our model

Industries · 7 min read

Fish size measurement on a grading line with a fixed camera

Fish boxed and sized against a calibrated channel, keypoints for true length when the tolerance demands it, and a bumped mount as the day every fish reads big.

Esdras Ntuyenabo · Sep 30, 2026

Bottling line, bottles queued under a fill head with one cap missing, bottle and cap boxed, generated scene with detections from our model

Industries · 6 min read

Object orientation from two keypoints, on a bottle line and a parcel sorter

Cap and base as keypoints, the angle between them as the answer, and consistent placement as the labeling rule that makes the whole thing hold.

Esdras Ntuyenabo · Sep 30, 2026

Camera on a pole over an industrial yard at dusk, vehicles and the fence boxed, generated scene with detections from our model

Industries · 6 min read

Occupancy analytics with computer vision for the lot, the dock and the floor

Each bay is a polygon, each sampled frame says occupied or free, and the whiteboard becomes a chart by hour. Snow is what the correction rate shows first.

Esdras Ntuyenabo · Sep 29, 2026

Orchard tree rows with three drivable paths marked from the tractor camera, from a customer tractor run

Computer vision · 6 min read

What semantic segmentation labels and what it costs

Drivable path, corrosion area and crop rows are questions about pixels, and the answer is a class for every one. Painted labels cost more than boxes.

Esdras Ntuyenabo · Sep 29, 2026

Loading dock from a mounted camera, a forklift with a pallet and trucks at the bays, generated scene with detections from our model

Labeling · 8 min read

Bounding boxes in computer vision, what a box teaches and what it reports

The same rectangle on a forklift is the lesson at training time and the answer at run time. Corner and centre formats, the shifted-box bug, floor in the box.

Esdras Ntuyenabo · Sep 28, 2026

Robotic welding cell with a part on the fixture, metal part and robot arm boxed, generated scene with detections from our model

Computer vision · 6 min read

When the outline is the answer and a box will not do

A weld defect whose area decides the rework and a part a robot has to grasp both need a mask. Here is what the mask costs to label and when a box is enough.

Esdras Ntuyenabo · Sep 28, 2026

Worker on a scaffold deck, hard hat and fall-arrest gear boxed, from a customer site camera

Labeling · 6 min read

Natural language image annotation, type what to look for and check what comes back

Hard hat and reflective vest, typed once and run across a site camera's frames. The boxes get checked one by one, and the lanyard clip gets drawn by hand.

Esdras Ntuyenabo · Sep 28, 2026

Crane lift from a pole camera, generated scene with detections from our model

Operations · 7 min read

Zone monitoring system built on a site camera and a polygon

A zone monitoring system from one site camera. A polygon over the crane swing radius, people and vehicles as classes, and the alert with the frame.

Esdras Ntuyenabo · Sep 27, 2026

Packing station from an overhead camera, boxes, a damaged box and a label boxed, generated scene with detections from our model

Operations · 7 min read

Event-driven object detection, turning a bounding box into an alert someone acts on

The packing area camera sees cartons all day. Three stacked is the rule, the frame goes to the supervisor, and a cooldown makes one stack one alert.

Esdras Ntuyenabo · Sep 27, 2026

Flat panels on a conveyor passing an inspection station, generated scene with detections from our model

Industries · 6 min read

Ceramic defect detection for hairline cracks a fixed-rule camera cannot learn

Edge chips, hairline cracks and pinholes on the tile line after the kiln, masks where the area sets the grade, and the new glaze as the day to relabel.

Esdras Ntuyenabo · Sep 26, 2026

Bottles queued under the fill head on a bottling line, generated scene with detections from our model

Industries · 6 min read

Expiration date detection on a curved bottle neck at the packing station

The code found as a box on the neck, read as a string, checked against the plant's own format, and the case held on a mismatch. A new coder is a spec change.

Esdras Ntuyenabo · Sep 26, 2026

Assembly station with a housing, a tray and a wire boxed, generated scene with detections from our model

Industries · 7 min read

PCB defect detection on bare boards, with a severity on every box

Opens, shorts and mouse bites boxed on the bare panel, critical ones rejected, repairable ones sent to rework with the frame, a new revision labeled early.

Esdras Ntuyenabo · Sep 26, 2026

Assembly bench with a housing, tray and wiring laid out, generated scene with detections from our model

Industries · 7 min read

Hog ring inspection on the seat line with object detection under the fabric

Every hog ring boxed and its closure read among wires and clips, the count checked against the build order, and the ring under the fabric is a labeling rule.

Esdras Ntuyenabo · Sep 25, 2026

Robotic weld cell with the torch over the fixture, generated scene with detections from our model

Industries · 7 min read

Robotic welding defect detection on every weld the robot lays down

A camera at the weld cell names porosity, undercut and missing fusion on every part, sends the borderline weld to a person, and a new torch needs new frames.

Esdras Ntuyenabo · Sep 25, 2026

Bottling line, one bottle missing its cap, generated scene with detections from our model

Industries · 7 min read

Bottle cap inspection with a camera at the capper outfeed

Sealed, loose and missing as three classes, a reject on the second two before the case packer, and the loose rate per batch as the number the technician wants.

Esdras Ntuyenabo · Sep 24, 2026

Stamping line, steel panels on a conveyor passing the inspection station, generated scene with detections from our model

Industries · 7 min read

How to detect metal defects with computer vision, one model across brushed and polished parts

Scratch, dent, crack and pit on a stamping line, the light that shows each, a decision layer that knows the cosmetic face, a new alloy labeled in advance.

Esdras Ntuyenabo · Sep 24, 2026

Packaging line, cartons with labels passing a scanner, carton and label boxed, generated scene with detections from our model

Industries · 7 min read

Label inspection with computer vision, six checks a carton passes before the case is sealed

Presence, placement, orientation, print, content and barcode, each check cropping the frame for the next, from the camera on the labeler outfeed.

Esdras Ntuyenabo · Sep 24, 2026

Bottling line with bottles queued under the fill head, bottle and cap boxed, generated scene with detections from our model

Industries · 7 min read

Glass defect detection where the conveyor shows through

Glass defect detection on a bottle line, where the conveyor shows through the glass and a scratch and a reflection arrive as the same bright line.

Esdras Ntuyenabo · Sep 23, 2026

Worker on a scaffold deck with hard hat and fall protection boxed, from a customer site camera

Computer vision · 5 min read

Pretrained vs fine-tuned detection models, when the stock detector cannot see the thing you need

A stock detector sees every worker on the scaffold and has no hard hat class, and no threshold will invent one. The site's own frames are what add it.

Esdras Ntuyenabo · Sep 23, 2026