Trash Car Train Data Set Preparation

[toc]

Introduction

There has 4 classes and 5 instance in the trash_c4_i5 dataset in total. This booklet is used to guide image taking and labeling to create trash_c4_i5 dataset.

Image prepare

1. Number We need 100 images per instance. 50 images on the green floor and 50 images on the mutative background.

2. Environment

   1. Green floor images:
      1. Multiple instances on the same images.
      2. Simulate real environment.
   2. Mutative background images: 3. Keep lighting strength changing. 4. Background changing. NO TWO SAME IMAGE.

3. Naming

   Image should start at 001.jpg and end at 100.jpg the corresponding label should as same as images.

4. Files Level

   trash_c4_i5
   |-- data
   	|-- predefined_classes.txt
   	|-- trash_c4_i5.yaml
   |-- images
       |-- train
           |-- 1.jpg
           ...
       |-- valid
           |-- 1.jpg
           ...
   |-- labels
       |-- train
           |-- 1.txt
           ...
       |-- valid
           |-- 1.txt
           ...
   

5. Camera and Image Size

   Use RGB camera which used in competition environment. And take the size of $640\times480$.

Label annotation

Label app: click to check and download.

The classes order is shown as below, DO NOT change this order.

bottle 
battery
cup
orange
paper

The file is arranged as below:

1. The default order is predefined in the file trash_c4_i5/data/predefined_classes.txt .

2. The annotations target dir is trash_c4_i5/labels/.

3. The images are saved in dir trash_c4_i5/images/.

The name of annotation must be as same as image. Must choose the yolo style to label all the images.

The useful hotkeys:

Ctrl + u	Load all of the images from a directory
Ctrl + r	Change the default annotation target dir
Ctrl + s	Save
Ctrl + d	Copy the current label and rect box
Ctrl + Shift + d	Delete the current image
Space	Flag the current image as verified
w	Create a rect box
d	Next image
a	Previous image
del	Delete the selected rect box
Ctrl++	Zoom in
Ctrl--	Zoom out
↑→↓←	Keyboard arrows to move selected rect box

Train

you can train with --img 1280 and/or --rect, though --rect is not recommended for best results. Which is trained as mosaics.

Reference

Image size

[1] ultralytics/yolov5#974

[2] ultralytics/yolov5#700

Multiple GPU:

[3] ultralytics/yolov5#475