pedropro/TACO

What it solves

TACO addresses the lack of diverse, real-world annotated images for training object detectors to recognize litter. It provides labeled photos of trash taken in varied environments such as woods, roads, and beaches, enabling more robust waste-detection systems.

How it works

Images are collected via an online server and manually labeled with segmentations following a hierarchical taxonomy. Annotations are supplied in COCO format for convenience, making them easy to use with common detection frameworks. The repository includes a download script, a Jupyter notebook for inspecting the data, and a modified Mask R-CNN detector with utilities to split data into train/val/test sets and to merge rare classes.

Who it's for

This is designed for researchers, engineers, and hobbyists working on computer vision, especially those focused on object detection for environmental monitoring, smart bins, or automated waste sorting. It also suits anyone needing a dataset to test or benchmark detection algorithms.

Highlights

• Growing image dataset of litter in natural and outdoor settings.
• Hierarchical taxonomy for detailed classification of waste types.
• Annotations provided in COCO format, a widely used standard.
• Includes a downloadable API, a demo notebook, and an active annotation server.
• Offers a Mask R-CNN detector implementation with customizable class mappings to handle sparse classes.