README.md

This dataset is created by NAVER Connect Foundation.

Recycle Trash Dataset

Overview

Recycle Trash Dataset is a dataset that collects indoor and outdoor garbage photos.
It consists of 21818 photos and annotation data for 107935 objects.
Each object has a bounding box and segmentation data and the compressed dataset is about 120 GB in size.

Contents

1. Labels
2. Samples
3. Dataset Structure
4. License

1. Labels

Each class is labeled according to standard recycling classification criteria.

Class Name	Object Count	Ratio
General trash	18117	16.79%
Metal	4451	4.12%
Plastic bag	23805	22.05%
Glass	4142	3.84%
Plastic	14098	13.06%
Styrofoam	5627	5.21%
Paper	29873	27.68%
Clothing	2091	1.94%
Paper pack	4235	3.92%
Battery	738	0.68%
UNKNOWN	758	0.7%
---	---	---
Total	107935	100%

2. Samples

General trash

Metal

Plastic bag

Glass

Plastic

Styrofoam

Paper

Clothing

Paper pack

Battery

3. Dataset Structure

1) Directory Structure

Each batch consists of 500 pictures and is randomly sampled over the entire data.
There is one json file in each batch and the annotation information is stored. 2 types of folder is provided.

1. data folder : original image size dataset for training.
2. vis (visualization) folder : vis folder is a visualization of the information in the json file and displays the bounding box and segmentation on the image that has been resized to 1080px in height. (only for preview purpose)

trash_dataset
├── data
│   ├── batch_01
│   │   ├── 0000.jpg
│   │   ├── 0001.jpg
│   │   ├── ...
│   │   ├── 0499.jpg
│   │   └── data.json
│   ├── batch_02
│   │   ├── 0000.jpg
│   │   ├── 0001.jpg
│   │   ├── ...
│   │   ├── 0499.jpg
│   │   └── data.json
│   ├── ...
│   ├── batch_44
│   │   ├── 0000.jpg
│   │   ├── 0001.jpg
│   │   ├── ...
│   │   ├── 0317.jpg
│   │   └── data.json

├── vis
│   ├── batch_01
│   │   ├── 0000.jpg
│   │   ├── 0001.jpg
│   │   ├── ...
│   │   ├── 0499.jpg
│   ├── batch_02
│   │   ├── 0000.jpg
│   │   ├── 0001.jpg
│   │   ├── ...
│   │   ├── 0499.jpg
│   ├── ...
│   ├── batch_44
│   │   ├── 0000.jpg
│   │   ├── 0001.jpg
│   │   ├── ...
│   │   ├── 0317.jpg

2) Json File Structure

For convenience, annotations are provided in COCO format.
Please check out the youtube video https://www.youtube.com/watch?v=h6s61a_pqfM

iscrowd variable is for surface masking using rle decode, but it is not used.

Segmentation points are in the form of x1, y1, x2, y2, x3, y3, ...,
and since one object may be divided into several, one object may be divided into several polygons.

4. License

This work is licensed under a Creative Commons Attribution 4.0 International License.