Object-Centric Scene Representation Learning via SAM

Recent advancements in the segmentation foundation model SAM have demonstrated exceptional performance in image segmentation. This study introduces object-centric representation learning, capitalizing on the segmentation outcomes derived from the foundational segmentation model SAM to elevate generalization performance and accelerate training within scene representation learning. Our approach involves a simple yet effective method to fine-tune unsupervised feature extraction in object-centric learning for scene representation. We conduct an effort to enhance the performance of unsupervised representation learning in out-of-distribution situations, resulting in improved performance in single-object out-of-distribution (OOD) scenarios.