Slot-oriented approaches for compositional scene segmentation from images and videos still depend on provided background information or slot assignments. We present Loci-Segmented (Loci-s) building on the slot-based location and identity tracking architecture Loci (Traub et al., ICLR 2023). Loci-s enables dynamic (i) background processing by means of a foreground identifying module and a background re-generator; (ii) top-down modified object-focused bottom-up processing; and (iii) depth estimate generation. We also improve automatic slot assignment via a slot-location-entity regularization mechanism and a prior segmentation network. The results reveal superior video decomposition performance in the MOVi datasets and in another established dataset collection targeting scene segmentation. Loci-s outperforms the state-of-the-art with respect to the intersection over union (IoU) score in the multi-object video dataset MOVi-E by a large margin and even without supervised slot assignments and without the provision of background information. We furthermore show that Loci-s generates well-interpretable latent representations. These representations may serve as a foundation-model-like interpretable basis for solving downstream tasks, such as grounding language, forming compositional rules, or solving one-shot reinforcement learning tasks.
翻译:暂无翻译