Vision Field Notes

Datasets and Training

15 articles

Datasets and Training

Writing an Annotation Guide Labellers Can Follow

Labels are only as consistent as the rules behind them. What to put in an annotation guide, how to show edge cases and how to keep it current.

5 min read
Datasets and Training

Annotation Mistakes That Quietly Ruin Your Model

Common detection annotation mistakes, from loose boxes to missing labels and inconsistent classes, and a practical way to audit and fix your dataset.

7 min read
Datasets and Training

Class Imbalance and Small Objects: Practical Fixes

Fixes for class imbalance and small objects in detection: more data, tiling, higher resolution, sampling and the right metrics to judge progress.

7 min read
Datasets and Training

CVAT vs Label Studio vs Roboflow: Choosing an Annotation Tool

CVAT vs Label Studio vs Roboflow compared on cost, hosting, export formats, team use and labelling speed, so you can pick an annotation tool.

7 min read
Datasets and Training

Data Augmentation for Object Detection: What Helps and What Hurts

Data augmentation for object detection explained: which transforms help, how mosaic and flips work, and how to avoid unrealistic ones that hurt.

6 min read
Datasets and Training

Deduplicating Video Frames Before Labelling

Most frames from a fixed camera are near-copies. How to pick frames worth labelling with difference checks, hashes and embeddings, keeping rare cases.

5 min read
Datasets and Training

Fine-Tuning YOLO: Transfer Learning Explained

Fine-tuning YOLO with transfer learning: why pretrained weights work, how to freeze layers, learning rate choices, and when fine-tuning fails.

7 min read
Datasets and Training

Hard Negative Mining from Production Footage

Turn the false alarms your cameras produce into training data: how to collect, review, label and balance hard negatives without breaking recall.

6 min read
Datasets and Training

How Many Images Do You Need to Train an Object Detector?

How many images to train an object detector? Realistic dataset sizes, how to know when to stop collecting, and ways to get more from a small dataset.

7 min read
Datasets and Training

Measuring Label Quality with Annotator Agreement

Double-label a sample, match the boxes and count disagreements. How to measure label quality, what it means for small objects, and what to do next.

5 min read
Datasets and Training

Pseudo-Labelling with a Larger Model: Time Saver or Data Poison?

Using a big model to pre-label images can cut annotation time, or quietly copy its mistakes into your dataset. How to use pseudo-labels safely.

5 min read
Datasets and Training

How to Spot Overfitting in Object Detection

How to spot overfitting in object detection: read loss curves and validation metrics, recognise the warning signs, and apply practical fixes that work.

6 min read
Datasets and Training

Train/Validation Split for Video Frames: Avoid Data Leakage

Why a random train/validation split of video frames inflates metrics, and how to split by video, camera and time so your detection results can be trusted.

6 min read
Datasets and Training

Train YOLO on a Custom Dataset: A Step-by-Step Guide

Train YOLO on a custom dataset with Ultralytics: folder layout, data.yaml, the training command, reading results and the first improvements to try.

7 min read
Datasets and Training

Why Your Test Set Flatters You

Test scores often come out higher than live performance. The usual causes are leakage, selective sampling, missing empty frames and a reused test set.

6 min read