WS2: Weakly Supervised Segmentation for Smarter Waste Sorting

The WS2 dataset is a benchmark dataset specifically designed for weakly supervised waste sorting segmentation. Collected from a plastic waste sorting facility, it comprises high-resolution video sequences captured before and after a human operator manually removes unwanted items from a stream of mixed plastic waste on a conveyor belt, leaving only semi-transparent colored PET items.

The dataset is designed to facilitate the segmentation of removed waste items by leveraging the implicit supervision provided by the operator. Specifically, before images contain both items to keep and items to remove, while after images include only the items to keep. This dataset enables the evaluation of weakly supervised methods (such as CAM-based approaches) in complex industrial environments, eliminating the need for extensive pixel-level annotations required for training fully supervised segmentation networks.

The dataset consists of a training set with 9,563 unannotated before-and-after images, already organized into training and validation folders, and a test set containing 1,497 before-and-after images with corresponding pixel-level ground truth masks. The test set masks are binary annotated, distinguishing items to be removed from the rest. Images are grouped into folders containing video sequences, allowing models to leverage temporal information and enabling the evaluation of both frame-based and video-based segmentation methods.

It includes both original images and background-removed versions, enabling direct comparisons between models trained with or without background information. The dataset follows a structured folder organization to support various training and evaluation setups.

• Full paper PDF:   WS2: Weakly Supervised Segmentation using Before-After Supervision in Waste Sorting.
• Experiments repo:   WS2: Weakly Supervised Waste Sorting Segmentation Benchmark.

If you use any ideas or materials from the paper, code or dataset from this repo, please consider citing:

@inproceedings{marelli2025ws2,
  author    = {Marelli, Andrea and Foresti, Alberto and Pesce, Leonardo and Boracchi, Giacomo and Grosso, Mario},
  title     = {WS2: Weakly Supervised Segmentation using Before-After Supervision in Waste Sorting},
  booktitle = {Computer Vision--ICCV 2025 Workshops},
  year      = {2025},
  url       = {https://arxiv.org/abs/2509.06485}
}