In traditional computer vision, we often ask models to perform a simple task: look at an image and pick the most important label. If a model sees a "wet dog," it succeeds by identifying that specific pair. However, the real world is rarely so sparse. A single object—like a "red, leather, vintage sneaker"—possesses a constellation of overlapping attributes.
Current research is shifting away from this "single-label" simplification toward Compositional Zero-Shot Learning (CZSL) that can handle multiple, concurrent properties. The goal is to build models that don't just recognize a "wet dog" because they've seen "wet grass," but can reason about an object's color, material, state, and texture all at once.
The Single Label Constraint
Most existing benchmarks for compositional learning, such as MIT-States or C-GQA, assign only one attribute-object pair to each image. As Muhammad Ferjad Naeem and colleagues point out in Learning Graph Embeddings for Compositional Zero-Shot Learning, this ignores the reality of natural settings. If a model predicts "brown dog" for an image labeled "wet dog," it is penalized, even if the dog is clearly both brown and wet.
This constraint creates a "bag-of-words" limitation. Models become optimized for specific pairs rather than understanding the underlying primitives. To solve this, researchers like Sheng Guo have introduced million-scale datasets like The iMaterialist Fashion Attribute Dataset, which force models to identify dozens of fine-grained attributes simultaneously, providing a more robust foundation than general-purpose datasets like ImageNet.
Disentangling Attributes from Objects
To recognize a "purple cube" without ever having seen one, a model must treat visual features as independent building blocks. This is known as disentanglement.
In Disentangling Visual Embeddings for Attributes and Objects, Nirat Saini and team proposed OADis, a framework that isolates attributes (like peeled) from objects (like apple) by comparing pairs of images. By "untangling" these features in the visual embedding space, the model can "imagine" what a never-before-seen combination looks like.
Similarly, Frank Ruis developed ProtoProp in Independent Prototype Propagation for Zero-Shot Compositionality. This method treats attributes as independent "prototypes" that propagate through a graph. By decoupling the "leather-ness" of a shoe from the shoe itself, the model can more easily apply that "leather" prototype to a new object, like a jacket.
Learning the Propagation Route
If attributes are independent, how do we know which ones belong together? Recent work focuses on how knowledge "flows" between concepts.
- Graph-Based Knowledge: Muhammad Ferjad Naeem introduced CGE, which builds a structured graph of states and objects. This allows the model to propagate information from seen combinations (wet grass) to unseen ones (wet dog) based on their structural relationships.
- Attention Propagation: In Learning Attention Propagation for Compositional Zero-Shot Learning, Muhammad Ferjad Naeem and Luc Van Gool proposed CAPE. Instead of static graphs, CAPE uses self-attention to discover dynamic "propagation routes" between compositions, helping the model understand hidden relationships between different object-state pairs.
Scaling to Real-World Complexity
The ultimate challenge is moving these models into "the wild," where multiple objects and textures coexist. Wei Zhai addresses this in texture recognition with MAP-Net, which learns to perceive multiple overlapping properties on a single surface—recognizing that a floor can be simultaneously "veined," "cracked," and "marbled" (Deep Multiple-Attribute-Perceived Network for Real-World Texture Recognition).
Finally, as we move toward autonomous systems, models must learn these compositions incrementally. Ran He and colleagues introduced CompILer in Not Just Object, But State: Compositional Incremental Learning without Forgetting. This approach uses specialized "prompt pools" to ensure that as a model learns new state-object pairs (like "yellow dress"), it doesn't forget the fine-grained attributes it learned previously.
By treating visual scenes as a multi-label puzzle rather than a single-label classification task, this research direction is enabling AI to capture the true density and complexity of the visual world.