Mikado Dataset

A Synthetic Dataset of Dense Homogeneous Object Layouts
for Occlusion-aware Instance Segmentation

Request access

This dataset is available only for research purposes. A small subset of the dataset can be downloaded here (55.2 MB). To acquire access to the whole dataset:

After your request is verified and approved by a website administrator, you will receive a notification by email. Please note, that any requests from free email addresses (hotmail, yahoo, gmail, etc.) will be refused.

Dataset content

The Mikado dataset is composed of 2,400 raw images, generated from 1 template mesh and 120 textures of sachets. Each scene consists of a top-view RGB image, an instance segmentation mask, and a distance map, as shown hereinafter. More details about the creation of the dataset can be found in the paper.

sample_image sample_instances sample_distances
RGB image Instance segmentation Distance map

The samples are structured by modality. Each downloadable zip file contains one modality.

All the modality directories have the same structure. Each modality directory contains directories, one for each textured object, which contain images. For example, if you want an RGB image along with the corresponding instances and boundaries, the relative paths to build are as follows:

Files format

For any sample, each modality is in PNG format: