YOLO

YOLO datasets that were measured, not annotated.

The YOLO format is deliberately minimal: a text file per image, five numbers per object. What it does not tell you is whether those five numbers are correct. Generated from an Unreal scene they come from the engine’s own identity buffer, so the box around a half-hidden crate is the box around the half you can see.

Classes: 6
Boxes in the sample: 1,813
Splits: train, val
Trainers: Ultralytics and compatible

The layout an Ultralytics trainer expects

Images and labels in parallel directory trees, matching stems, and a data.yaml at the root. That is all, and getting it slightly wrong is the most common reason a first training run reports zero objects.

dataset/
├── data.yaml
├── images/
│   ├── train/   31 frames
│   └── val/     9 frames
└── labels/
    ├── train/   one .txt per image, same stem
    └── val/

# data.yaml
path: .
train: images/train
val: images/val
test: images/test

nc: 6
names:
  0: person
  1: car
  2: tank
  3: container
  4: crate
  5: barrel

What is in a label file

One line per object: class id, then centre x, centre y, width and height as fractions of the image size. Below is the start of a real exported file, copied verbatim.

# labels/train/dump__plugin_000004.txt
5 0.63125000 0.82546296 0.06041667 0.08055556
3 0.57317708 0.83287037 0.04739583 0.06944444
4 0.66718750 0.92685185 0.02604167 0.04629630
4 0.69609375 0.91388889 0.05260417 0.10555556

Read the first line: class 5 is a barrel, sitting at 63% across and 83% down the frame, occupying 6% of the width. On a 1920×1080 image that is a 116 by 87 pixel box, near the bottom of the frame, which is exactly where the barrels are.

What the export does that a naive one does not

Four decisions that separate a dataset you can train on from a directory of text files.

Boxes come from pixels

Not from projected 3D bounds. An object mostly behind a hangar gets a box around the visible part, because the identity buffer only contains the part that was drawn.

Degenerate boxes are dropped

Anything under 8 pixels on a side, or with an aspect ratio beyond 6:1, is rejected before export rather than exported as an unlearnable target.

Splits are checked for leakage

Consecutive frames of one capture look nearly identical. The validator hashes images and reports near-duplicates across the train and validation split. The reference export reported zero.

Class ids are stable

Declaration order, written into data.yaml and used consistently. They do not shuffle between runs because a class happened to appear in a different order.

# export YOLO only, or both formats from one capture
nameframe dataset  _out/dump _out/dataset --format yolo --val-frac 0.2
nameframe validate _out/dataset --out _out/validation

# then train
yolo detect train data=_out/dataset/data.yaml model=yolo11n.pt epochs=100 imgsz=1024

Before your first epoch

Two things to look at in the export, both of which are cheaper to notice now.

Class balance. The reference capture has eighteen barrels for every tank, because the scene does. Check the measured balance and decide whether to weight the loss or spawn differently.
Object size. More than half of the instances in an aerial capture are small by COCO’s definition. If your image size is 640, most of them are gone before the first convolution.

The sample pack contains the images, the labels, the data.yaml, the COCO version of the same annotations, and the report that graded all of it. Download it (49 MB) and check both before writing any code.