Ensemble Foreground Management for Unsupervised Object Discovery

Unsupervised object discovery (UOD) aims to detect and segment objects in 2D images without handcrafted anno- tations. Recent progress in self-supervised representation learning [9, 66] has led to some success in UOD algo- rithms [21, 53, 69]. However, the absence of ground truth provides existing UOD methods with two challenges: 1) determining if a discovered region is foreground or back- ground, and 2) knowing how many objects remain undiscov- ered. To address these two problems, previous solutions rely on foreground priors [53, 59, 67, 69] to distinguish if the discovered region is foreground, and conduct one or fixed it- erations of discovery. However, the existing foreground pri- ors are heuristic and not always robust, and a fixed number of discoveries leads to under or over-segmentation, since the number of objects in images varies. This paper intro- duces UnionCut, a robust and well-grounded foreground prior based on min-cut [2] and ensemble methods [18] that detects the union of foreground areas of an image, allow- ing UOD algorithms to identify foreground objects and stop discovery once the majority of the foreground union in the image is segmented. In addition, we propose UnionSeg, a distilled transformer of UnionCut that outputs the fore- ground union more efficiently and accurately. Our experiments show that by combining with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] wit- ness an increase in the performance of single object discovery, saliency detection and self-supervised instance seg- mentation on various benchmarks. The code is available at https://github.com/YFaris/UnionCut.