co-FCN(Conditional Networks for Few-Shot Semantic Segmentation)[1] handle sparse pixel-wise annotations to achieve nearly the same accuracy. There are some details of reading and implementing it.

Contents

• Paper & Code & note
• Paper understanding
• Code understanding
• Note
• References

Paper & Code & note

Paper: Conditional Networks for Few-Shot Semantic Segmentation(ICLR 2018 paper)
Code: [Code]
Note: Mendeley

Paper

Abstract

1. They propose the co-FCN, a conditional network learned by end-to-end optimization to perform fast, accurate few-shot segmentation.
2. The network conditions on an annotated support set of images via feature fusion to do inference on an unannotated query image.
3. Annotations are instead conditioned on in a single forward pass, making our method suitable for interactive use.

Problem Description

Some current methods rely on meta-learning, or learning to learn, in order to quickly adapt to new domains or tasks.

• It cannot be applied out-of-the-box to the structured output setting due to the high dimensionality of the output space.
• The statistical dependencies among outputs that result from the spatial correlation of pixels in the input.

Problem Solution

• Our contributions cover handling sparse pixel-wise annotations, conditioning features vs. parameters.
• Our method achieves nearly the same accuracy with only one positive and one negative pixel.

Conceptual Understanding

Experiments

Code

[Updating]

Note

[Updating]

References

[1] Rakelly K, Shelhamer E, Darrell T, et al. Conditional networks for few-shot semantic segmentation[J]. 2018.