pedropro/TACO

Provides instance segmentation annotations in COCO format across a hierarchical waste taxonomy, with images sourced from diverse real-world environments (woods, roads, beaches). Includes a Mask R-CNN detector implementation with configurable class mappings to handle severe class imbalance, plus Python utilities for dataset splitting, downloading from Flickr, and inspecting annotations via Jupyter notebooks. Supports both official curated annotations and community-submitted unofficial annotations tracked separately for quality control.

713 stars. No commits in the last 6 months.