Complete Guide to Creating COCO Datasets
4.6 (221 ratings)
Course Ratings are calculated from individual students’ ratings and a variety of other signals, like age of rating and reliability, to ensure that they reflect course quality fairly and accurately.
541 students enrolled

Complete Guide to Creating COCO Datasets

Build your own image datasets automatically with Python
4.6 (221 ratings)
Course Ratings are calculated from individual students’ ratings and a variety of other signals, like age of rating and reliability, to ensure that they reflect course quality fairly and accurately.
541 students enrolled
Last updated 4/2019
English
English [Auto]
Price: $39.99
30-Day Money-Back Guarantee
This course includes
  • 4 hours on-demand video
  • 1 article
  • 6 downloadable resources
  • Full lifetime access
  • Access on mobile and TV
  • Certificate of Completion
Training 5 or more people?

Get your team access to 4,000+ top Udemy courses anytime, anywhere.

Try Udemy for Business
What you'll learn
  • How COCO annotations work and how to parse them with Python
  • How to go beyond the original 90 categories of the COCO dataset
  • How to automatically generate a huge synthetic COCO dataset with instance annotations
  • How to train a Mask R-CNN to detect your own custom object categories in real photos
Course content
Expand all 53 lectures 03:55:56
+ COCO Image Viewer
10 lectures 43:55
Initialization
04:11
Processing COCO Instances JSON
04:57
Display info, licenses, and categories
02:54
Display Image: Open and calculate resize ratio
04:04
Display Image: Polygon segmentations
08:16
Display Image: RLE segmentation concept
05:20
Display Image: RLE segmentation code
08:39
Running the notebook on the COCO Dataset
02:27

A few questions to make sure you understood this section.

COCO Image Viewer Quiz
2 questions
+ Dataset Creation with GIMP
7 lectures 35:40
Section Introduction
01:24
Opening, scaling, and exporting
07:16
Create first mask and export
06:41
Use layers to create second mask
04:54

Soundtrack by Silk Music: http://www.youtube.com/SilkMusic
Track name: LTN - Thunderball

Remaining images time-lapse
02:05
Mask definitions JSON
06:09
Mask definitions JSON (remaining images)
07:11
GIMP Dataset Creation Quiz
3 questions
+ COCO JSON Utils
9 lectures 41:41
Section Introduction
00:32
Context for coco_json_utils.py
01:23
Overview
06:04
Validate and process arguments and create info
04:37
Create licenses and categories
04:35
Create images and annotations
06:22
Split multicolored mask into isolated masks
05:54
Create annotations with isolated masks
09:29
Running coco_json_utils.py
02:45
COCO JSON Utils Quiz
1 question
+ Foreground Cutouts with GIMP
5 lectures 19:18
Section Introduction
01:24
Context for cutting out foregrounds
02:55
Foreground Select Tool (rough)
06:35
Foreground Select Tool (clean) and export
06:21
Free Select Tool with Feather Edges
02:03
Foreground Cutouts Quiz
2 questions
+ Image Composition
12 lectures 01:00:41
Section Introduction
01:24
MaskJsonUtils overview, adding categories, and adding masks
08:05
Getting masks, getting super categories, and writing to json
03:31
ImageComposition overview
05:26
Validate and process arguments
04:16
Validate and process foregrounds and backgrounds
04:28
Choose random foregrounds and background
03:40
Crop background and transform foregrounds
06:07
Compose images and masks
08:03
Save images and mask definitions json
05:21
Create dataset info
03:12
Running image_composition.py and coco_json_utils.py
07:08
Image Composition Quiz
2 questions
+ Training Mask R-CNN
5 lectures 21:38
Section Introduction
00:41
Getting started with Mask R-CNN
06:48
Preparing to train with our synthetic dataset
05:13
Training
04:17
Running inference on real test images
04:39
+ Course Wrap
1 lecture 01:10
Overview and wrap
01:10
Last Quiz (It's easy, I promise)
1 question
Requirements
  • Take an Intro to Deep Learning course first (perhaps from another Udemy course!)
  • Have basic to intermediate Python programming skills
  • Have a physical or cloud computer with GPU/CUDA compute
  • Recommended: Prior experience with Anaconda & Jupyter notebooks
Description

In this course, you'll learn how to create your own COCO dataset with images containing custom object categories. You'll learn how to use the GIMP image editor and Python code to automatically generate thousands of realistic, synthetic images with minimal manual effort. I'll walk you through all of the code, which is available on GitHub, so that you can understand it at a fundamental level and modify it for your own needs.

(Important: If you only want to do manual image annotation, this course is not for you. Google "coco annotator" for a great tool you can use. This course teaches how to generate datasets automatically.)

By the end of this course, you will:

  • Have a full understanding of how COCO datasets work

  • Know how to use GIMP to create the components that go into a synthetic image dataset

  • Understand how to use code to generate COCO Instances Annotations in JSON format

  • Create your own custom training dataset with thousands of images, automatically

  • Train a Mask R-CNN to spot and mark the exact pixels of custom object categories

  • Be able to apply this knowledge to real world problems

I've saved weeks of my precious time using this method because I'm not doing the tedious task of manual image labeling, which can easily take a full 40 hour work week to create 1000 images. You should value your time too. After all, how are you going to solve the world's problems if you're busy clicking outlines on images for the next couple weeks?


Soundtrack by Silk Music
Track name: Shingo Nakamura - Hakodate

Who this course is for:
  • Developers who have completed a Deep Learning course and want to solve real-world image recognition problems
  • Developers looking for a deep walkthrough of creating a COCO dataset and training a Mask R-CNN