Rapidata/coco-human-inpainted-objects

About:

The dataset was collected on the https://www.rapidata.ai platform and contains tens of thousands of human annotations of 70+ different kinds of objects. Rapidata makes it easy to collect manual labels in several data modalities with this repository containing freehand drawings on ~2000 images from the COCO dataset. Users are shown an image and are asked to paint a class of objects with a brush tool - there is always a single such object on the image, so the task is not ambiguous. The result of this user-interaction is a collection of lines drawn by the user on that particular image.

Challenge:

The challenge of the dataset is to aggregate the lines on each image to get an idea of where the target object is. For each image, we provide hundreds of 2D lines drawn by different humans that can be used to create bounding boxes and segmentation maps on each image of the target object. Apart from the lines, the dataset contains the COCO 2D bounding box ground truths as well as baseline predictions to beat.

Structure:

The metadata.csv describes each image in one row:

Column Name	Description
coco_filename	The unique identifier for each image in the COCO dataset.
class_name	The class/category that the user was asked to mark on the image. Same as category_name in COCO.
prediction	A baseline COCO bounding box prediction based on heatmaps.
ground_truth	The COCO bounding box ground truth.
IoU	The Intersection over Union (IoU) score between the baseline prediction and the ground truth.
lines	A 3D array of coordinates. Because each user can draw multiple lines, the first dimension represents different users, the second dimension represents multiple lines drawn by each user, and the third dimension represents the individual [x, y] coordinates of each line, relative to the image dimensions.