Skip to content

Dataset formats

Argus supports YOLO, COCO, and folder-based semantic mask datasets. Detection and segmentation are handled out of the box.

YOLO

Argus looks for a YAML config file with a names key. It uses that file to extract class names and verify the dataset layout.

Typical structure:

dataset/
├── data.yaml
├── images/
│   ├── train/
│   ├── val/
│   └── test/
└── labels/
    ├── train/
    ├── val/
    └── test/

Unsplit YOLO datasets are also supported:

dataset/
├── data.yaml
├── images/
└── labels/

Roboflow YOLO

Argus also supports the Roboflow variant of the YOLO format, where images and labels live inside split directories instead of under a top-level images/ and labels/ folder:

dataset/
├── data.yaml
├── train/
│   ├── images/
│   └── labels/
├── valid/
│   ├── images/
│   └── labels/
└── test/
    ├── images/
    └── labels/

Roboflow uses valid instead of val for the validation split. Argus detects this automatically and normalises it to val internally. The data.yaml paths typically use ../train/images style references — Argus handles these by falling back to filesystem probing when the relative paths do not resolve.

Argus infers the task type by scanning a few label files:

  • 5 values per line: detection
  • more than 5 values per line: segmentation polygons

COCO

Argus looks for COCO annotation JSON files in annotations/ or at the dataset root.

Typical structure:

dataset/
├── annotations/
│   ├── instances_train.json
│   ├── instances_val.json
│   └── instances_test.json
└── images/
    ├── train/
    ├── val/
    └── test/

If your annotation filenames include train, val, or test, Argus will treat those as splits. Otherwise the dataset is treated as unsplit.

Roboflow COCO

Argus also supports the Roboflow variant of COCO format, where annotations live inside split directories:

dataset/
├── train/
│   ├── _annotations.coco.json
│   └── *.jpg
├── valid/
│   ├── _annotations.coco.json
│   └── *.jpg
└── test/
    ├── _annotations.coco.json
    └── *.jpg

Splits are detected from directory names (train, valid/val, test).

COCO RLE segmentation

COCO annotations can encode instance masks using Run-Length Encoding (RLE) instead of polygon point lists. Argus detects RLE annotations automatically and renders them as pixel-level mask overlays in the viewer.

Both RLE variants are supported:

  • Uncompressed RLE — counts is a list of integers: 
    {
  "segmentation": {
    "counts": [0, 5, 10, 3, 82],
    "size": [100, 100]
  }
}
  • Compressed RLE — counts is a COCO-style LEB128 string: 
    {
  "segmentation": {
    "counts": "eNpjYBgFo2AU0AsABEAAEQ==",
    "size": [100, 100]
  }
}

Datasets that mix RLE and polygon annotations in the same file are handled correctly. When you run argus view on a COCO dataset with RLE annotations, Argus opens the mask overlay viewer instead of the bounding-box viewer.

Mask (semantic segmentation)

Mask datasets are simple image + mask folders. Argus detects a few common patterns:

  • images/ + masks/
  • img/ + gt/
  • leftImg8bit/ + gtFine/ (Cityscapes-style)

Split-aware layout:

dataset/
├── images/
│   ├── train/
│   └── val/
├── masks/
│   ├── train/
│   └── val/
└── classes.yaml  # Optional for grayscale, required for RGB palette masks

Unsplit layout:

dataset/
├── images/
├── masks/
└── classes.yaml

Mask encoding

  • Grayscale masks: each pixel value is the class ID. Argus will auto-detect class IDs if no classes.yaml is provided.
  • RGB palette masks: each class maps to a color. A classes.yaml is required.
  • Mask files should be .png and match the image stem (e.g., frame.png), or use common suffixes like _mask, _gt, or _label.

Example classes.yaml:

names:
  - background
  - road
  - sidewalk
ignore_index: 255
palette:
  - id: 0
    name: background
  - id: 1
    name: road

Detection heuristics

If Argus does not detect your dataset, check the following:

  • The dataset root is correct and readable.
  • YOLO: the YAML file includes names as a list or dict.
  • COCO: images, annotations, and categories exist in the JSON.