Author

Stephen Biswas

Engineer · LexData · 35 articles

Two workers under safety netting on a construction site camera, hard hats boxed and the plank decks marked, from a customer site camera

Computer vision · 6 min read

What pose estimation gives a site camera that a box cannot

A box says a worker is there. Keypoints say the worker is bending, reaching, or has a wrist a metre from a live terminal, and that is the safety question.

Stephen Biswas · Oct 4, 2026

Factory floor, a robot arm loading parts into a CNC machine behind a safety fence, fence and robot arm boxed, generated scene with detections from our model

Computer vision · 7 min read

YOLO training best practices, what decides whether the model survives the factory floor

Frames from the floor itself, a labeling spec written before the first box, per-class recall over the average, and a loop that keeps running after go-live.

Stephen Biswas · Oct 4, 2026

Thermal camera on a substation fence line at night, people at the fence, vehicles and the fence line boxed, from a customer site camera

Labeling · 7 min read

Brightness and contrast adjustment for annotation, labeling the defect the camera barely shows

A faintly hot bushing on a night thermal frame is there and the labeler cannot see it. Turn the display up, draw the box, and leave the training pixels alone.

Stephen Biswas · Oct 3, 2026

Shopper at an open dairy case, partly behind the door, rack sections and two carts boxed, from a customer store camera

Labeling · 7 min read

Occlusion in computer vision, training for the half the camera can see

A forklift behind a rack, a shopper behind a display, a part under the robot arm. On a fixed camera the partial view is the normal one, and the labels say so.

Stephen Biswas · Oct 3, 2026

Drone pass along a solar farm, panels boxed with straight boxes across the rows, from a customer drone survey

Computer vision · 6 min read

Oriented bounding box detection, when the box needs to turn with the object

A straight box on a panel tilted at forty-five degrees is mostly background, and on a packed array the boxes overlap. A rotated box fixes that, at a price.

Stephen Biswas · Oct 3, 2026

Dashcam frame with traffic signs, pedestrians and vehicles boxed, from a customer perception run

Labeling · 7 min read

Stop sign detection edge cases and the long tail of a simple object

A stop sign is the easiest class there is until it is snowed on, held by a crossing guard or printed on a bus. The tail is covered one doubted frame at a time.

Stephen Biswas · Oct 3, 2026

Pipeline across scrub, cradles boxed and a rust mark flagged, from a customer drone survey

Operations · 6 min read

Batch video analysis of archived drone survey footage without a notebook open

Three seasons of right-of-way flights in a cloud bucket. Import the originals, sample the frames, run the model, and review only what it doubted.

Stephen Biswas · Oct 2, 2026

Stamping line, steel sheets passing the inspection station, generated scene with detections from our model

Labeling · 7 min read

Data augmentation for object detection on a fixed inspection camera

A stamping line camera varies in known ways. Conveyor speed changes scale and blur, a relamp changes the light. Augment for those and move the boxes too.

Stephen Biswas · Oct 2, 2026

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

Labeling · 6 min read

Image preprocessing vs data augmentation, and why only one runs at the camera

Resize, orientation and normalisation have to match between the labeled frames and the runner. Augmentation runs in training only, never at the camera.

Stephen Biswas · Oct 2, 2026

Pipe rack with rust marked along five pipe sections, from a customer inspection run

Labeling · 7 min read

SAM assisted segmentation labeling, click the patch and correct the mask

A labeler clicks a corroded patch, a mask appears, and the job becomes correcting an edge instead of tracing one. The person earns it where the proposal fails.

Stephen Biswas · Oct 2, 2026

Assembly station under a fixed camera, generated scene with detections from our model

Operations · 6 min read

How to choose a camera for computer vision before you collect a frame

Work back from the smallest thing the model must see, then shutter, lamp and interface. A camera swapped later is a sensor replacement with a relabel.

Stephen Biswas · Oct 1, 2026

Bottles queued under a fill head with one cap missing, generated scene with detections from our model

Operations · 6 min read

Industrial machine vision lighting and optical filters decide what the camera sees before any model does

The scratch an inspector finds by tilting the part under a window is invisible under the wrong lamp. Light it once, and treat a lamp change as a new camera.

Stephen Biswas · Oct 1, 2026

Steel panels on a stamping line passing an inspection station, generated scene with detections from our model

Operations · 6 min read

Camera calibration for computer vision, and why the part at the edge of the frame measures wrong

A straight edge bows at the corner of the frame, so a part that passes in the centre fails at the edge. Calibrate once, and again the day the lens changes.

Stephen Biswas · Sep 30, 2026

Parcel packing station from overhead, cartons, a crushed carton and a label boxed, generated scene with detections from our model

Industries · 7 min read

Object dimension measurement from a fixed camera, using the mask rather than the bounding box

A rotated carton needs a polygon, scale comes from a printed reference or depth, the carrier's bands decide pass or fail, and a nudged mount breaks it.

Stephen Biswas · Sep 30, 2026

Packaging line, cartons on the conveyor past a scanner, generated scene with detections from our model

Industries · 7 min read

Product colour inspection with a camera, measured inside the mask

A mask keeps the belt out, the colour is quantised to swatches and compared with a reference measured on the same camera, and a swapped camera moves everything.

Stephen Biswas · Sep 30, 2026

Pole transformer photographed from the ground, transformer and insulators boxed, from a customer inspection run

Labeling · 7 min read

What is EXIF orientation and why training sees it sideways

EXIF orientation is a tag telling a viewer to rotate the picture while the pixels stay put, so a phone photo looks upright and arrives sideways.

Stephen Biswas · Sep 29, 2026

Checkout lane from a ceiling camera, a person and the products on the belt boxed, generated scene with detections from our model

Computer vision · 6 min read

Why a good detector can still be a bad counter

The queue camera finds every shopper on every frame and still gets the count wrong. Swaps, lost tracks and jitter are tracker mistakes, and review fixes one.

Stephen Biswas · Sep 29, 2026

Vehicle door sill and wheel boxed in a claims photo, generated scene with detections from our model

Industries · 6 min read

Tire sidewall OCR, reading the DOT code off a black tire in a dark bay

Box the sidewall, light it from the side, read the DOT date code and size from the crop, and log it against the vehicle. A swapped bay camera is a new sensor.

Stephen Biswas · Sep 29, 2026

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

Labeling · 7 min read

How to label image data for computer vision so the model learns what you meant

Tight boxes on the forklift, the visible part of a hidden pallet, a written rule for the frame edge, and why a forgotten box teaches the model it is background.

Stephen Biswas · Sep 28, 2026

Thermal camera on a fence line at night, two people at the fence and the vehicles boxed, from a customer site camera

Computer vision · 6 min read

What mAP measures and what it cannot tell you about your cameras

Mean average precision is built from per-class curves at an overlap line. A high score on the held-out set says nothing about the fence camera at 2 am.

Stephen Biswas · Sep 28, 2026

Food processing line, fillets on a blue conveyor with the conveyor and rail boxed, generated scene with detections from our model

Computer vision · 6 min read

ResNet-18 explained, what a residual backbone does underneath your detector

Why the shortcut connection made deep networks trainable, where the eighteen layers sit inside a detector on a plant camera, and why the frames matter more.

Stephen Biswas · Sep 28, 2026

Shopper at an open dairy case, rack sections and carts boxed, from a customer store camera

Operations · 8 min read

Dwell time detection in video, measuring how long rather than how many

A shopper at the dairy case, a worker inside a guard, a truck at a dock. Dwell needs identity across frames, a zone with a soft edge, and stream timestamps.

Stephen Biswas · Sep 27, 2026

Warehouse aisle with a forklift and racked pallets from a high camera, generated scene with detections from our model

Operations · 6 min read

Presence and absence detection, is the barrier in place and how long has it been gone

A forklift bay, a shadow board and a safety gate as present or absent per frame, a rule that fires after four minutes gone, and the pallet in the way.

Stephen Biswas · Sep 27, 2026

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

Industries · 6 min read

Chocolate box inspection with a camera over the tray line

Each piece boxed by type, a plain check against the box template, damage as its own class, and the new spring assortment as the day the labels go stale.

Stephen Biswas · Sep 26, 2026

Cartons passing a scanner on a packaging line, carton and conveyor boxed, generated scene with detections from our model

Industries · 7 min read

Lot code and expiry date verification with vision AI before the pouch moves on

The strip found on every pouch wherever the film pulled it, the crop read, format and date logic checked, and a failed read stopping the pouch, not guessing.

Stephen Biswas · Sep 26, 2026

Cartons passing a scanner on a packaging line, generated scene with detections from our model

Industries · 6 min read

Optical character verification checks the print against the work order instead of reading it from scratch

The expected lot and expiry come from the order, a detector finds the print on every pack, and the pack is held when they disagree. A new layout is a relabel.

Stephen Biswas · Sep 26, 2026

Analog gauge on a plant pipe with the needle boxed, generated scene with detections from our model

Industries · 8 min read

Computer vision water meter reading, a camera in the vault

Computer vision water meter reading turns a truck roll into a frame. Ten digit classes in a row, and a rolling digit or fogged glass goes back to a person.

Stephen Biswas · Sep 25, 2026

Bottling line with bottles queued under a fill head, generated scene with detections from our model

Industries · 7 min read

Leak inspection of IV bags with a line camera

Leak inspection of IV bags by camera, with fill level from the same frame. Fluid and ports find the bag, and a wet port is a question put to the frame.

Stephen Biswas · Sep 25, 2026

Transmission tower from a drone pass, insulator strings boxed, from a customer inspection run

Industries · 7 min read

Object detection on aerial imagery and why the ground detector fails at four hundred feet

Insulators a few dozen pixels wide, pipelines at every angle, solar tables packed edge to edge, and why a re-planned flight path is a camera that moved.

Stephen Biswas · Sep 25, 2026

Fillets on a food processing conveyor, generated scene with detections from our model

Industries · 6 min read

Coffee bean inspection with computer vision on the roasted bean belt

A camera over the roasted bean conveyor finds stones, quakers and foreign material as masks. A new origin lot is the day the defect rate moves.

Stephen Biswas · Sep 24, 2026

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

Industries · 7 min read

Container yard management with computer vision, knowing where every box sits without a radio call

Row cameras and a hostler camera box each container and its ID panel, vote the number across a pass, and answer where the reefer went last night.

Stephen Biswas · Sep 24, 2026

Assembly station with a housing and a cable loom laid out, generated scene with detections from our model

Industries · 6 min read

Object alignment detection for the label, the connector and the gasket on one station

Is it where it should be, and turned the right way. A box answers most of it, pose answers the last two degrees, and the reference moves when the camera does.

Stephen Biswas · Sep 24, 2026

Edge of a solar farm on the visible-light pass, panels boxed on the nearest table, from a customer drone survey

Industries · 7 min read

Thermal drone solar panel hotspot detection, and the palette that fools the model

Hotspots as boxes against a uniform module, white-hot against ironbow as what changes the pixels, the thermographer confirming, a new camera as a new baseline.

Stephen Biswas · Sep 24, 2026

Assembly station with the tool, tray and loom wire boxed, generated scene with detections from our model

Operations · 6 min read

Turning a detection into an OPC UA tag the PLC can read

A count, a presence flag and a confidence value as OPC UA variables, written on change from the alert webhook, while the reject decision stays with the PLC.

Stephen Biswas · Sep 23, 2026

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

Industries · 6 min read

Choosing a machine vision system for visual inspection on the line

Backlight, ring, coaxial or dome for the defect you need to see, why a badly lit frame cannot be fixed later, and vendor questions about the second year.

Stephen Biswas · Sep 23, 2026