metadata

license: mit
task_categories:
  - object-detection
size_categories:
  - 10K<n<100K
tags:
  - coco
  - bodyparts
  - body-parts

COCO Human Parts Dataset

This is a subset of the COCO dataset specifically designed for human body part detection. The dataset includes detailed annotations for each image, allowing for training models to not just detect humans but to also identify and localize specific parts of the human body.

Labels

The COCO Human Parts dataset contains the following labels:

person [0]
head [1]
face [2]
lefthand [3]
righthand [4]
leftfoot [5]
rightfoot [6]

These labels represent the different body parts that the dataset focuses on. Each body part is annotated with a bounding box in the YOLO format (x_center, y_center, width, height). The bounding boxes are normalized such that their coordinates range from 0 to 1 relative to the image's dimensions.

Data Format

The annotations are saved in Ultralytics' YOLO format. In this format, each image is associated with a text file of the same name. This text file contains one line for each bounding box in the image, with five numbers on each line separated by spaces. The five numbers represent the object class (one of the labels above), along with the centerX, centerY, width, and height of the bounding box. All the values are relative to the width and height of the image itself, ranging from 0 to 1.

Data Splits

The dataset is split into two sets:

Train Set: Contains 64115 images.
Val Set: Contains 2693 images.

Data Statistics

Data statistics for the COCO Human Parts dataset have been generated from logs produced during training with Ultralytics.

Label Distribution

The following plot shows the label distribution in the dataset, providing insight into the frequency of each body part being annotated.

Sample Images

Below are some sample images from the train set, along with the bounding box annotations for each body part. These images have been augmented using data augmentation techniques in Ultralytics training logs.

Citation

Images and annotations are from the COCO dataset.

Humanbodayparts annotations are from Hier-R-CNN.