dataset_info:
features:
- name: id
dtype: string
- name: split
dtype: string
- name: image
dtype: image
- name: seg
dtype: image
- name: num_classes
dtype: int64
- name: inds
sequence:
sequence: int64
- name: ids
sequence: string
- name: classes
sequence: string
- name: sizes
sequence:
sequence: float64
- name: centers
sequence:
sequence:
sequence: float64
- name: occlusions
sequence:
sequence: int64
- name: area
sequence:
sequence: int64
- name: unoccluded_area
sequence:
sequence: int64
- name: bboxes
sequence:
sequence:
sequence:
sequence: int64
- name: centers_crop672
sequence:
sequence:
sequence: float64
- name: occlusions_crop672
sequence:
sequence: int64
- name: area_crop672
sequence:
sequence: int64
- name: bboxes_crop672
sequence:
sequence:
sequence:
sequence: int64
splits:
- name: val
num_bytes: 2755568627.75
num_examples: 2413
- name: test
num_bytes: 2361310977.25
num_examples: 2115
- name: train
num_bytes: 5265670004.75
num_examples: 4757
download_size: 10255201099
dataset_size: 10382549609.75
configs:
- config_name: default
data_files:
- split: val
path: data/val-*
- split: test
path: data/test-*
- split: train
path: data/train-*
license: mit
Multi-Class Class-Agnostic Counting Dataset
**Project Page | ArXiv | Download **
Michael Hobley, Victor Adrian Prisacariu.
Active Vision Lab (AVL), University of Oxford.
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.
File Hierarchy
File Hierarchy:
├── dataset_pytorch.py
├── make_gaussian_maps.py
├── test
├── train
│ ├── 1511489148409439
│ ├── 3527550462177290
│ | ├──img.png
│ | ├──info.json
│ | ├──seg.png
│ ├──4109417696451021
│ └── ...
└── val
Precompute Density Maps
To precompute ground truth density maps for other resolutions, occlusion percentages, and gaussian standard deviations use the code from our GitHub:
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>;
Pytorch Dataset
There is a pytorch dataset written on our GitHub. This randomises the bounding boxes durig training but uses consistent bounding boxes for testing.
Citation
@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},
}