**Project Page | ArXiv | Download **

Michael Hobley, Victor Adrian Prisacariu.

Active Vision Lab (AVL), University of Oxford.

Each object in the RGB image has an associated: Model ID, Class ID, Center Coordinate, Bounding Box and Occlusion.

MCAC is the first multi-class class-agnostic counting dataset. each image contains between 1 and 4 classes of object and between 1 and 300 objects per class. The classes of objects present in the Train, Test and Val splits are mutually exclusive, and where possible aligned with the class splits in FSC-133. Each object is labeled with an instance, class and model number as well as its center coordinate, bounding box coordinates and its percentage occlusion Models are taken from [ShapeNetSem]. The original model IDs and manually verified category labels are preserved. MCAC-M1 is the single-class images from MCAC. This is useful when comparing methods that are not suited to multi-class cases.

Dowload MCAC here.

To precompute ground truth density maps for other resolutions, occlusion percentages, and gaussian standard deviations:

cd PATH/TO/MCAC/
python make_gaussian_maps.py  --occulsion_limit <desired_max_occlusion>  --crop_size 672 --img_size <desired_resolution> --gauss_constant <desired_gaussian_std>;

@article{hobley2023abc,
    title={ABC Easy as 123: A Blind Counter for Exemplar-Free Multi-Class Class-agnostic Counting}, 
    author={Michael A. Hobley and Victor A. Prisacariu},
    journal={arXiv preprint arXiv:2309.04820},
    year={2023},
}