Unsupervised Instance Segmentation with Superpixels (arxiv.org)

Researchers introduced an unsupervised instance segmentation framework that produces high-quality object masks without any human annotations. The pipeline first applies a MultiCut algorithm to self-supervised image features to generate coarse masks, then filters those masks to retain only high-quality candidates. Superpixels computed from low-level image cues are integrated into training via a novel superpixel-guided mask loss composed of a hard loss (discrete agreement) and a soft loss (probabilistic consistency). Finally, a self-training stage with an adaptive loss refines predicted masks and improves robustness. This approach is significant because it reduces dependency on costly pixel-level labels while still improving instance segmentation and object detection performance: the authors report outperforming previous unsupervised state-of-the-art methods on public benchmarks. Key technical implications include leveraging combinatorial segmentation (MultiCut) over self-supervised embeddings, using superpixels as structural priors to impose local consistency during learning, and employing an adaptive loss in self-training to gradually trust model predictions. The design is particularly promising for robotics, autonomous systems, and large-scale vision datasets where annotation is infeasible, offering a scalable route to instance-level understanding with competitive accuracy.