hwanji/taco

🌮 Trash Annotations in Context Dataset Toolkit

TACO is a growing image dataset of waste in the wild. It contains images of litter taken under diverse environments: woods, roads and beaches. These images are manually labeled and segmented according to a hierarchical taxonomy to train and evaluate object detection algorithms. Currently, images are hosted on Flickr and we have a server that is collecting more images and annotations @ tacodataset.org

For convenience, annotations are provided in COCO format. Check the metadata here: http://cocodataset.org/#format-data

TACO is still relatively small, but it is growing. Stay tuned!

Publications

For more details check our paper: https://arxiv.org/abs/2003.06975

If you use this dataset and API in a publication, please cite us using:  

@article{taco2020,
    title={TACO: Trash Annotations in Context for Litter Detection},
    author={Pedro F Proença and Pedro Simões},
    journal={arXiv preprint arXiv:2003.06975},
    year={2020}
}

News

December 20, 2019 - Added more 785 images and 2642 litter segmentations.
November 20, 2019 - TACO is officially open for new annotations: http://tacodataset.org/annotate

Getting started

Requirements

To install the required python packages simply type

pip3 install -r requirements.txt

Additionaly, to use demo.pynb, you will also need coco python api. You can get this using

pip3 install git+https://github.com/philferriere/cocoapi.git

Download

To download the dataset images simply issue

python3 download.py

Alternatively, download from

Our API contains a jupyter notebook demo.pynb to inspect the dataset and visualize annotations.

Trash Detection

The implementation of Mask R-CNN by Matterport is included in /detector with a few modifications. Requirements are the same. Before using this, the dataset needs to be split. You can either donwload our weights and splits or generate these from scratch using the split_dataset.py script to generate N random train, val, test subsets. For example, run this inside the directory detector:

python3 split_dataset.py --dataset_dir ../data

For further usage instructions, check detector/detector.py.

As you can see here, most of the original classes of TACO have very few annotations, therefore these must be either left out or merged together. Depending on the problem, detector/taco_config contains several class maps to target classes, which maintain the most dominant classes, e.g., Can, Bottles and Plastic bags. Feel free to make your own classes.