Related Research On Perceptual Grouping Problems

Learning effective distributed and disentangled representation is an important research direction of unsupervised learning.There is ambiguity problem,called as binding problem,in the unbinding recognition of the scene generated by the mixture of multiple effective representations.The binding problem widely exists in real-world applications of computer vision and speech,such as scene segmentation in autonomous driving and virtual reality,and multi-speaker speech segmentation,etc.Perceptual Grouping(PG)is an important mechanism to solve the binding problem,it has the ability to learn the complete and effective representation of each entity object from the complex structured input scene.Therefore,the research on perceptual grouping related problems has crucial practical significance.In the existing unsupervised visual perceptual grouping algorithms,for the complex scenes composed by multiple overlapping entity objects,the results of perceptual grouping present a serious problem of fuzzy and incomplete representation of entity object reconstructions,and when the entity objects in the scene are grouped at pixel level,the phenomenon of unreasonable grouping is serious.In this paper,based on the principles of spatial mixture and adversarial learning,an unsupervised Adversarial Perceptual Grouping(APG)algorithm is proposed.APG combines the advantages of the reasoning model,which can map the real input to the hidden variable space,and the adversarial network,a high-quality generation model.On the basis of the spatial mixture model,Generative Modeling of Infinite Occluded Objects for Compositional Scene Representation(GMIOOCSR)which is suitable for complex scenes with multiple overlapping entity objects,APG introduces the discriminator in an adversarial learning way to implicitly learn the rich information about Gestalt laws of grouping among multiple observations of the entity objects.The part of spatial mixture model learns the effective representation of the entity objects,autonomously determines the existence of the entity objects in the scene,and performs pixel-level grouping based on the learned representation and existence of the entity objects,and reconstructs the scene according to the hierarchical structure among objects.The reconstruction error of the discriminator is used to further constrain the spatial mixture model,so as to make the results of perceptual grouping model more complete,real and reasonable.APG is compared with two state-of-the-art algorithms on benchmark datasets Multi-Shapes and Multi-MNIST.The quantitative comparative experiment using three perceptual grouping performance evaluation indicators,the visual qualitative comparative experiment,the effect experiment of feature validity and the comparative experiment of generalization ability in three new scene settings are carried out.Experimental results show that the entity object representations learned by APG have a good distributed and disentangled property,and APG achieves better performance of perceptual grouping and stronger generalization ability,indicating that the introduction of adversarial structure effectively enhances the perceptual grouping effect on complex image scenes.