A scene is not only a field of pixels. It is often a collection of things that persist, move, interact and can be rearranged. Object-centric representation tries to make that compositional structure explicit.
Instead of representing an image or world state only as one distributed vector, object-centric models learn a set of latent slots or entities, each intended to bind to a distinct component of the scene. Those objects can then carry attributes, identities and relations across time.
This article extends the eduKateSingapore World Representation & Cognitive Tools branch into compositional latent structure: how a model represents a world that contains several persistent parts.
The Object-Centric Route
SCENE / VIDEO / SENSOR INPUT → PERCEPTUAL FEATURES → COMPETITIVE BINDING → OBJECT SLOTS → ATTRIBUTES + POSITION + IDENTITY → RELATIONS BETWEEN OBJECTS → TEMPORAL PERSISTENCE → COMPOSABLE WORLD STATE → REASONING / PREDICTION / CONTROL
1. Object-Centric Representation Is Not Ordinary Object Detection
Object detection predicts labels and boxes from a predefined or supervised ontology. Object-centric representation asks a broader question: can the model discover reusable entity-like latent units that explain scene structure even without explicit object labels?
The goal is a compositional internal representation, not merely a final detection output.
2. Objects Are Natural Units for Recombination
A red cube can move from left to right while remaining the same cube. A chair can appear in many rooms. The same relation “on top of” can connect different pairs of objects.
Entity-like representations support this kind of reusable structure.
3. Slot Attention Learns a Set of Object-Like Latents
Slot Attention introduces a competitive attention mechanism that maps perceptual features into a set of exchangeable latent slots. Each slot can specialise to part of the input, producing abstract object-centric representations.
The slots are not assigned permanent meanings in advance.
4. Slots Are Exchangeable
Slot 1 does not have to mean “car” forever and Slot 2 “person”.
The slots behave like a set: different slots can bind to different objects on each input while the overall representation remains permutation-aware rather than sequence-dependent.
5. Competition Encourages Specialisation
Input features compete for assignment across slots rather than every slot independently attending to everything.
This exclusivity pressure helps different slots capture different scene components.
6. Competition Does Not Guarantee One Slot per Human Object
A slot can split one object, merge several objects or bind to background.
Object-centricity is an emergent representational property that needs evaluation, not an automatic consequence of using slots.
7. Binding Is the Core Problem
A scene contains features such as colour, shape, position and motion. The model must bind the right features together as belonging to the same entity.
Without binding, “red”, “moving right” and “cube” can become detached facts.
8. Object Slots Are Variable Containers
A slot can encode several attributes simultaneously: appearance, geometry, category, pose or hidden state.
This makes each slot a compact local world model for one entity.
9. Attributes Should Be Disentangled Only When the Task Needs It
Explicitly separating colour, shape and position can support compositional transfer.
But forcing every factor into a clean human axis can reduce flexibility if real factors interact.
10. Pose Should Transform Predictably
An object moved across the image should ideally retain identity while its position component changes.
This motivates equivariant object representations.
11. Invariant Slot Attention Builds Object-Centric Reference Frames
Invariant Slot Attention introduces slot-centric reference frames that translate, scale and rotate positional information to improve object discovery under spatial transformations.
The broader principle is that objects should preserve identity while their pose transforms systematically.
12. Object Identity Across Time Is Harder Than Object Discovery in One Frame
A slot that represents a person in one frame may switch to another person in the next.
Persistent world models need identity continuity, not only frame-level segmentation.
13. Identity Requires Temporal Binding
Motion, appearance, predicted position and interaction history can all support tracking.
No single cue is always sufficient.
14. Occlusion Is a Persistence Test
When an object disappears behind another object, a strong world model should continue to represent its probable existence and state.
Object permanence requires latent memory beyond immediate visibility.
15. Reappearance Tests Identity Recovery
When the hidden object becomes visible again, the system should connect the new observation to the previously tracked entity rather than create a new identity casually.
This joins object-centric representation to memory and belief state.
16. Objects Become More Powerful When Relations Are Explicit
A scene is not just a bag of isolated entities. Objects can be above, behind, touching, supporting, approaching or owned by other entities.
Relational structure can be represented as edges among object slots.
17. Object Sets Naturally Become Graphs
Slots provide nodes; learned or observed relations provide edges.
This allows graph reasoning while preserving object identity.
18. Relation Types Can Be Learned or Symbolic
A model can learn continuous relational features or map them to named predicates such as “inside” or “connected_to”.
Continuous relations are expressive; symbolic relations are easier to audit.
19. Object-Centric Models Support Compositional Generalisation
If object identity and relation are factorised, a model can encounter familiar objects in unfamiliar combinations.
The original Slot Attention work demonstrated improved generalisation to unseen compositions in controlled tasks.
20. Composition Is Stronger Than Memorising Whole Scenes
A flat scene embedding can memorise common layouts. An object-relational representation can reuse parts under new arrangements.
This is one reason object-centricity is attractive for planning and reasoning.
21. Object Slots Can Feed World Models
Instead of predicting one monolithic latent state, a dynamics model can predict how each object changes and how interactions couple those changes.
See World-State Representation.
22. Object-Centric Dynamics Can Be More Modular
Moving one object need not require rewriting the entire scene representation.
Local intervention can produce local state change plus relational consequences.
23. Modular Dynamics Connect to Causal Representation
Objects are plausible units for intervention because many actions target particular entities.
But object-centricity does not itself prove causal factorisation. See Causal Representation Learning.
24. Parts and Objects Create Hierarchy
A car contains wheels; a person contains limbs; a room contains furniture.
One slot level may not be enough for complex compositional scenes.
25. Hierarchical Object-Centric Representation Can Model Part–Whole Structure
Lower-level slots can represent parts while higher-level slots represent composed entities.
This connects object-centricity to hierarchical representation.
26. Slot Count Is a Capacity Budget
If the model has fewer slots than objects, several entities may be merged. If it has many extra slots, some can remain empty or fragment the scene.
Slot count therefore encodes an assumption about scene complexity.
27. Fixed Slot Count Is Convenient and Restrictive
Real scenes contain variable numbers of entities.
Dynamic slot allocation is a natural extension when object count varies widely.
28. Empty Slots Need a Meaningful Null State
A slot with no object should not accidentally bind to background noise and later be interpreted as a real entity.
Presence or confidence is part of slot state.
29. Background Is a Representation Choice
Some models allocate one slot to background; others distribute background across slots.
The decision changes how foreground object boundaries emerge.
30. Object Boundaries Are Not Always Visually Sharp
Smoke, shadows, water, cloth and crowds challenge the assumption that scenes decompose into clean rigid objects.
Object-centric models should not force discrete entities where the world is better represented as fields or collective structure.
31. Event-Centric Representation Can Complement Objects
Some tasks care more about interactions and transformations than persistent entities.
An advanced world model may combine object state with event state.
32. Language Can Name Object Slots
Visual object representations can be aligned with words or phrases such as “red mug” or “left door”.
Language adds human-readable identity while the latent slot retains perceptual state.
33. Naming Can Improve Interface and Introduce Bias
A language label can make a slot auditable. It can also force perceptual variation into an imperfect human category.
Symbol grounding should preserve both the label and underlying evidence.
34. Identifiability Is an Advanced Object-Centric Problem
Several different slot decompositions can reconstruct the same scene.
Recent work on probabilistic Slot Attention studies conditions under which object-centric slots can be identified up to an equivalence class rather than arbitrarily.
35. Reconstruction Alone Does Not Guarantee the Right Decomposition
A model can reconstruct an image accurately while grouping pixels into latent parts that do not correspond to stable entities.
Object quality needs independent tests.
36. Object Discovery Evaluation Needs More Than Segmentation Score
Test identity persistence, compositional generalisation, relation prediction, intervention response and downstream planning.
A pretty mask is not a sufficient world model.
37. Planning Is a Strong Stress Test
If a robot must move one object around another, the representation must preserve identity, position, collision geometry and relational consequences.
Object slots that drift or swap identities can break planning even when frame reconstruction remains good.
38. Object-Centric Representation Audit
- What counts as an entity for this receiver job?
- How are slots initialised and updated?
- What competitive mechanism encourages separate binding?
- Does one slot usually correspond to one meaningful entity?
- How are empty slots represented?
- How does slot count relate to object count?
- Can identity persist across motion and occlusion?
- Are pose changes represented equivariantly?
- What attributes live inside each slot?
- How are relations between slots represented?
- Can familiar objects recombine in unseen arrangements?
- Does the representation support action-conditioned prediction?
- What happens for non-object-like phenomena such as fluids or crowds?
- Can outputs return to the source pixels, frames or sensor evidence supporting each object?
39. What Students Should Remember
- Object-centric models represent scenes as sets of entity-like latent units.
- Slot Attention learns exchangeable slots through competitive attention.
- Binding groups attributes that belong to the same object.
- Temporal persistence requires identity tracking beyond one frame.
- Object relations naturally form graph structure.
- Compositional representations can generalise to unseen arrangements.
- Object-centricity helps world modelling but does not automatically prove causality.
40. The Deep Principle
The world often changes by moving and transforming parts rather than redrawing itself from scratch.
Object-centric representation is powerful when it turns perception into persistent, recombinable entities whose identities survive change, whose relations can be reasoned over, and whose latent claims can still be traced back to the scene that produced them.