Simple to Complex Curriculum for Weakly Supervised Vision

This direction is merged into another direction: curriculum learning. Its last version stays readable and citable here.

Version 1 of 124 papers35 conceptsfrozen 2026-10-04

Weakly supervised learning presents a classic "chicken-and-egg" problem: to find an object in an image, a model needs to know what it looks like; but to learn what it looks like from image-level tags alone, the model must first find it. When training on complex, cluttered scenes, models often fail by latching onto "co-occurring" background noise—associating the label "boat" with water rather than the vessel itself.

The research direction emerging from these papers solves this by adopting a Simple-to-Complex (STC) curriculum. By sequencing training data so that a model masters "easy" images before tackling "hard" ones, researchers have found they can bridge the gap between global image labels and pixel-level understanding.

Bootstrapping from Visual Prototypes

The intuition behind this approach is that not all training samples are created equal. Some images are "cleaner" than others. Yunchao Wei and colleagues in STC: A Simple to Complex Framework for Weakly-Supervised Semantic Segmentation argue that images containing a single, centered object against a clean background act as a natural starting point.

In these simple cases, the ambiguity of the image-level label is minimized. If a label says "dog" and there is only one prominent object in the center, the model can easily learn a reliable "visual prototype" of a dog. Qinbin Hou and Daniela Massiceti demonstrated that initializing models on these simplified datasets provides a massive performance jump—up to 10%—because it prevents the model from getting lost in the noise of complex scenes during the early stages of training.

Defining the Curriculum

How do we mathematically define what makes an image "simple"? Researchers have proposed several quantitative metrics to rank training difficulty:

  • Object Count and Category Density: Dingwen Zhang and Deyu Meng suggest a category-count curriculum. In Bridging Saliency Detection to Weakly Supervised Object Detection, they show that starting with images containing only one object category allows the model to establish reliable boundaries before it has to resolve the spatial overlap found in multi-category scenes.
  • Object Size: In medical imaging, Gui Wang and Yuexiang Li use size as a proxy for difficulty. Their S³-Mamba framework trains on large, clear lesions first to build a robust feature foundation before attempting to segment tiny, subtle lesions that standard models often overlook.
  • Linguistic and Semantic Complexity: For multimodal models, Jiacheng Ruan and Zehao Lin use the length of instructions and image-text embedding similarity. In their work on MM-CamObj, they show that mastering basic visual-textual associations first is essential for eventually detecting highly deceptive, camouflaged objects.

The Zigzag Path to Robustness

While starting simple is effective, there is a risk that the model will overfit to the most "discriminative" parts of an object—learning only the head of a cat rather than its whole body.

To counter this, Xiaopeng Zhang and Jiashi Feng introduced Zigzag Learning in Zigzag Learning for Weakly Supervised Object Detection. This approach alternates between two strategies: it progressively introduces harder images while simultaneously masking out the most obvious parts of the objects it has already found. This forces the model to look for less obvious features, ensuring that the "simple" lessons translate into a comprehensive understanding of the entire object.

Why It Works: Establishing an Appearance Prior

The success of these methods, as surveyed by Dingwen Zhang in Weakly Supervised Object Localization and Detection: A Survey, lies in the establishment of a strong appearance prior.

In a standard weakly supervised setting, the model is bombarded with ambiguity. By the time it sees a cluttered image of a "park," it might have already decided that "park" means "green grass." By using a curriculum, the model learns what the individual components (trees, benches, dogs) look like in isolation. When it finally encounters a complex scene, it doesn't have to learn from scratch; it simply has to localize the patterns it already recognizes. This staged approach effectively "denoises" the weak supervision, allowing models to achieve results that rival fully supervised systems with significantly less human labeling effort.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_simple_to_complex_curriculum_for_weakly__d19a5d83,
  title = {Simple to Complex Curriculum for Weakly Supervised Vision},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simple-to-complex-curriculum-for-weakly-supervised-vision/rdr_d19a5d83a8f64d4d96094b776bebc709},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_simple_to_complex_curriculum_for_weakly__d19a5d83_v1,
  title = {Simple to Complex Curriculum for Weakly Supervised Vision},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simple-to-complex-curriculum-for-weakly-supervised-vision/rdr_d19a5d83a8f64d4d96094b776bebc709/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}