---

To view the dataset, visit https://seattletrashproject.com

Seattle Trash Project (STP) is an open-source vision dataset of post-consumer waste identified in Seattle. The dataset is hosted in S3/Dynamo (AWS) and is available for public scaled consumption.

Note: Please be courteous when downloading data from the dataset. The dataset is provided publicly at no cost but hosting and accessing the dataset is not free. Transfers are collectively throttled so downloading multiple times might block someone else from downloading the dataset.

Download

Manual

The dataset is hosted in S3 & Dynamo. The S3 bucket with dataset images is public and can be accessed via web or with AWS SDKs. Annotations are hosted in Dynamo and can be accessed via API Gateway endpoints.

S3 Bucket: recyclr-dataset

• Contains image .png files. Each object key is named with the md5 sum of the raw image.

Coco Dataset Generator: https://prw168zo7a.execute-api.us-west-2.amazonaws.com/prod/coco

• Creates a coco json file that can be used to train or evaluate models.

Automated

First, install dependencies:

# pip install boto3 requests

To download the raw dataset, run:

# python download_stp.py

This will download annotations.json containing a coco-compatible dataset file. This will also download the raw images (.png) from the dataset into the ./images subdirectory. As of 11/2021 the raw dataset is ~4.86GB.

To download the downsized dataset, run:

# python download_stp.py --small

This will download annotations.json containing a coco-comaptible dataset file. This will also download the miniaturized image (.png) for each element from the dataset into the ./images-small subdirectory. As of 11/2021, the small dataset is ~424MB.

Format

Example annotation:

{
    "segmentation": [
        [
            1769.1874322860226,
            2300.671722643552,
            1463.4019501625125,
            2624.6587215601285,
            1448.8407367280597,
            2608.2773564463687,
            1754.6262188515698,
            2282.470205850486
        ]
    ],
    "bbox": [
        1448.8407367280597,
        2282.470205850486,
        320.3466955579629,
        342.1885157096426
    ],
    "image_id": 0,
    "category_id": 54,
    "id": 0,
    "iscrowd": 0,
    "material": "polypropylene",
    "element": "straw",
    "mass": 2.0,
    "quality": 0.87,
    "producer": "mcdonalds"
}

Each annotation in the dataset contains four keys:

Key	Description	Example
material	the dominant material the object is made of	polypropylene, aluminum, fiber, etc.
element	the type (class) of object	beverage bottle, cigarette, bottlecap, etc.
mass	the mass (g) of an object in grams	15.0
quality	rough probability a post-consumer item can be recovered [0-1.0]	0.59
producer	(optional) producer associated with the item	coca cola, starbucks, etc.

The coco category map defaults to element. annotations.json will include a categoryMap object which maps class names to category_id, ex:

categoryMap = {
    "beverage bottle" : 1,
    "bottlecap" : 2,
    ...
}