Stats and counts
argus-cv stats provides per-class instance counts and image totals by split.
For mask datasets, it reports pixel coverage and how many images contain each
class.
Example
Argus prints a table by class and split. It also includes a summary with:
- total instances
- number of classes
- image totals and background images
Why background counts matter
Empty label files or images without annotations can skew training. Argus counts those so you can decide if you want to filter or re-label.
Roboflow YOLO datasets
Roboflow exports YOLO datasets with a different directory layout
({split}/images/ instead of images/{split}/). Argus detects this
automatically — no extra flags needed. The valid directory is treated as the
val split.
Common problems
If Argus prints "No annotations found", check:
- YOLO:
labels/exists and matchesimages/. For Roboflow YOLO datasets, labels should be at{split}/labels/. - COCO: annotation JSON files are valid and contain
annotations. - Mask: masks are
.pngfiles and match the image file names.