Filtering datasets

Use argus-cv filter to create a filtered copy of a dataset containing only specified classes.

Relative output paths are resolved under the dataset root. For example, argus-cv filter /datasets/coco -o filtered --classes person writes to /datasets/coco/filtered.

Basic usage

argus-cv filter /datasets/coco -o /datasets/coco_filtered --classes person,car

This creates a new dataset with only the person and car classes. Class IDs are automatically remapped to sequential values (0, 1, 2, ...).

Filter to a single class

argus-cv filter /datasets/yolo -o /datasets/yolo_balls --classes ball

Exclude background images

By default, images without annotations (after filtering) are kept. Use --no-background to exclude them:

argus-cv filter /datasets/coco -o /datasets/coco_filtered --classes dog --no-background

This is useful when you want a dataset with only images that contain your target class.

Use symlinks for faster filtering

For large datasets, use --symlinks to create symbolic links instead of copying images:

argus-cv filter /datasets/large -o /datasets/filtered --classes cat --symlinks

This saves disk space and speeds up the filtering process significantly.

Supported formats

The filter command works with all dataset formats:

Format	Supported	Notes
YOLO Detection	Yes	Labels remapped to new class IDs
YOLO Segmentation	Yes	Polygon annotations preserved
YOLO Classification	Yes	Only selected class directories copied
YOLO (Roboflow layout)	Yes	Reads `{split}/images/` layout, writes standard layout
COCO	Yes	Annotations and category IDs remapped
Mask	Yes	Pixel values remapped to new class IDs

Output layout

The output always uses the standard dataset layout, even when the source uses a different directory structure (e.g. Roboflow YOLO). This normalises the output so it can be used directly with training frameworks.

YOLO output:

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

COCO output:

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

Class ID remapping

When filtering, class IDs are remapped to start from 0 and be sequential. For example:

Original	Filtered
0: person	(removed)
1: car	0: car
2: dog	1: dog
3: cat	(removed)

If you filter to keep only car and dog, the new dataset will have car as class 0 and dog as class 1.

Common errors

"No classes specified": You must provide at least one class name with --classes.
"Classes not found in dataset": Check the class names match exactly (case-sensitive). Use argus-cv stats to see available classes.
"Output directory already exists": The output directory must be empty or non-existent.