puje123 / miu_ml

Resources for MIU ML course

MIU Machine Learning

This repository is for Machine Learning class at MIU. Unless the students choose a different project, the default project is "the Mongolian Sidewalk Project" described below.

Mongolian Sidewalk Project

The goal of the Mongolian Sidewalk Project is build an AI that detects various objects seen in the streets of Mongolia. By doing the project, the students will have a chance to see the whole aspects of developing an AI and thereby learn Machine Learning.

What needs to be done?

Here is what is expected for all participants.

1. Collecting data: Take pictures of the streets in Mongolia and add to "data/sidewalk_mn/to_label" folder.
2. Label: Use makesense.ai to label target objects (See Dataset section for details). We are using bounding box label type. After you are done with labeling, export the annotation in YOLO format.
3. Train: As a starter, we will use YOLO V5.
4. Predict: After the model is trained, we will use a testset to evaluate the model performance
5. Deploy: For a demo purpose, we will create a simple web service to predict real data.

Dataset

To train a model, we need a lot of data. Our target goal is to have at least 300 labeled images.

Each participant is expected to contribute at least 50 labeled images. After 300 labeled images are gathered, we will split the data into:

• train: 200 images
• validation: 50 images
• test: 50 images

To annotate

After the images are collected, you need to label them. You can use makesense. Upload sidewalk_labels.txt as labels file. After having with annotation, export the labels in YOLO format and check them into the data folder.

Classes

The initial dataset classes are borrowed from the Korean sidewalk dataset. There might be different objects we are interested in identifying so we will leave the option to add more classes later on. The initial classes are:

id	class label
0	wheelchair
1	truck
2	tree_trunk
3	traffic_sign
4	traffic_light
5	traffic_light_controller
6	table
7	stroller
8	stop
9	scooter
10	potted_plant
11	power_controller
12	pole
13	person
14	parking_meter
15	movable_signage
16	motorcycle
17	kiosk
18	fire_hydrant
19	dog
20	chair
21	cat
22	carrier
23	car
24	bus
25	bollard
26	bicycle
27	bench
28	barricade

Train

There are many object detection models. As a starter, we can use YOLO V5 which is one of the most commonly used object detection models. You can take a look at sidewalk_ko_yolov5 to see how training is done. We want to replicate it for Mongolian data and build a model that can detect object in Mongolian streets.

Related Work

There have been several Machine Learning projects that deal with sidewalk image data. For instance, Deep Learning for Automatically Detecting Sidewalk Accessibility Problems Using Streetscape Imagery (2019), and Korean sidewalk project for people with disabilities.