LinSlot
A paper introducing the Linear Representation Hypothesis (LRH) into slot-based visual object representation. Using a probabilistic graphical model, it achieves disentangled representations of objects and their attributes in an unsupervised manner, enabling controllable image editing without supervision.
Timeline (last 90 days)
LinSlot introduces the Linear Representation Hypothesis (LRH) into the slot-based visual object representation framework.
LinSlot constructs a probabilistic graphical model connecting images, slots (objects), and blocks (attributes), and is trained by optimizing the Evidence Lower Bound (ELBO).
The linearly additive subspaces of LinSlot provide two benefits—disentanglement and editability—enabling modification of object attributes through subspace addition and subtraction operations.
LinSlot surpasses the current state-of-the-art methods on the DCI disentanglement quality score.
LinSlot can discover disentangled object and attribute representations without any manual annotation and supports image editing.
LinSlot has currently been validated primarily on relatively structured datasets, and the robustness of LRH in more complex natural scenes remains to be examined.
All Facts (6)
LinSlot introduces the Linear Representation Hypothesis (LRH) into the slot-based visual object representation framework.
50%UnverifiedLinSlot constructs a probabilistic graphical model connecting images, slots (objects), and blocks (attributes), and is trained by optimizing the Evidence Lower Bound (ELBO).
50%UnverifiedThe linearly additive subspaces of LinSlot provide two benefits—disentanglement and editability—enabling modification of object attributes through subspace addition and subtraction operations.
50%UnverifiedLinSlot surpasses the current state-of-the-art methods on the DCI disentanglement quality score.
50%UnverifiedLinSlot can discover disentangled object and attribute representations without any manual annotation and supports image editing.
50%UnverifiedLinSlot has currently been validated primarily on relatively structured datasets, and the robustness of LRH in more complex natural scenes remains to be examined.
50%