Keywords: object-centric learning, structured representations, unsupervised learning, deep learning
TL;DR: We perform a systematic experimental study on the separation performance of object-centric models with different inductive biases on datasets with objects that have complex texture.
Abstract: Understanding which inductive biases could be useful for unsupervised models to learn object-centric representations of natural scenes is challenging. Here, we use neural style transfer to generate datasets where objects have complex textures. Our main finding is that a model that reconstructs both the shape and visual appearance of each object from its representation achieves correct separation of the objects and learns useful object representations more reliably.
3 Replies
Loading