Computer Vision Case Studies

18 case studies

Faster blood cell labeling with live training

Label a few blood smears by hand and let a model that learns as you go prelabel the rest. Before long, your team just reviews.

Labeling
Object Detection
Label
Train
Image
Medical Imaging

Dataset audit before training: what is wrong with PCB Component Detection

One run audits a PCB detection dataset for label errors, near-duplicate images, split leakage and outliers, ranks the findings and says what to fix first.

Data Quality
Inspect
Evaluate
Image
Industrial Inspection

Supervisely labeling queues: one shared pool, nobody waiting

Annotators take the next image the moment they are free instead of waiting on a batch, and every image's stage stays visible on the canvas.

Labeling
Label
Review
Image
Supervisely

Auto quality check: catch mistakes before review

Catch annotation errors with rule-based validation before datasets reach manual review.

Labeling
Label
Review
Image
Industrial Inspection

Labeling jobs in Supervisely, run by a flow

Keep labeling in Supervisely and let a flow cut the batches, create the jobs, re-issue rejected work, and show where every image is right now.

Labeling
Label
Review
Image
Supervisely

Consensus review: several opinions where one is not enough

Every annotation is judged by a panel instead of one reviewer, and only a genuine disagreement reaches an expert — so the specialist spends time on the hard cases, not on the whole batch.

Labeling
Label
Review
Image
Medical Imaging

Labeling with an expert for the hard cases

Label X-rays with a team, escalate the hard cases to an expert, and review every annotation before it counts, without moving a single file yourself.

Labeling
Label
Review
Image
Medical Imaging

Dataset cleanup: turn statistics into fixes

Use dataset profiling to find the problems in a dataset and fix them before training starts, instead of discovering them after a run that already cost you GPU hours.

Data Quality
Evaluate
Clean
Image
Industrial Inspection

Basic labeling pipeline: a label-and-review loop for your team

How a team labels and reviews images in the base OvalBee labeling loop. A universal starting point you can customize for any project.

Labeling
Label
Review
Image

CLIP: anomaly alerts for every new batch of data

Index your data with CLIP, describe the images you don't want in plain words, and split each batch into clean data and anomalies — with an email when a batch loses more than it should.

Data Quality
Dataset Curation
Search
Clean
Clip
Image
Traffic

How to reduce duplicate video frames before labeling

Select the most diverse driving scenes from the comma2k19 driving video dataset. Maximize visual diversity and minimize near-duplicates while preserving broad, representative coverage of the original dataset.

Dataset Curation
Sample
Diversify
Clip
Video

Dataset inspection

A dataset quality report turns a folder of images into a dozen charts. What each one is for, what to look for in it, and the kind of defect it exposes - from records that hold nothing to a class name typed wrong.

Data Quality
Evaluate
Image
Industrial Inspection

How to use CLIP for diverse dataset curation

AI Search narrows a road-scene dataset with CLIP text search and diverse sampling; Similar pairs exposes remaining repetition for review.

Dataset Curation
Search
Diversify
Inspect
Clip
Image

Three detectors, one scoring pass: does the scoreboard hold up?

Three off-the-shelf detectors scored against the same COCO val2017 reference by the same evaluation pass, checked for the things that make an evaluation trustworthy before it is pointed at private data.

Object Detection
Evaluate
Compare
Image
Yolo
Deim
Edgecrafter
Public Benchmark

32.1% mAP against traffic ground truth — and three reasons the number is not the model

One pretrained detector scored against a 226-image labeled traffic set, where the per-class breakdown says something very different from the headline aggregates — and two classes score zero for a reason that has nothing to do with detection.

Object Detection
Evaluate
Image
Edgecrafter
Traffic

Did it converge? Fine-tuning a detector on PCB component photos

One run takes 1,099 labeled circuit-board images to a validated EdgeCrafter checkpoint at 68.1% mAP, with the per-epoch curve that shows where training stopped paying.

Object Detection
Train
Image
Edgecrafter
Industrial Inspection

Before you label: does an off-the-shelf detector see your objects?

One pass runs a pretrained EdgeCrafter detector over 226 unlabeled traffic images, and the returned boxes answer two questions before any annotation budget is spent — does the model react, and in whose vocabulary.

Object Detection
Inference
Image
Edgecrafter
Traffic

Two architectures, one PCB dataset: which detector do you ship?

EdgeCrafter and DEIM are trained and evaluated on the same PCB component data, turning a model choice into a repeatable, evidence-backed decision.

Object Detection
Train
Compare
Image
Deim
Edgecrafter
Industrial Inspection