Article directory
- foreword
- overview
- Proposed benchmark
- Overall Mechanism with CD-FSS
- patnet
- Task-Adaptive Fine-Tuned Inference (TFI) Strategy
- Experiment and Visualize
- Summarize
- reference link
- Articles in the same series
foreword
Following the medical image processing series, we have returned to the topic of semantic segmentation of small samples. I also sorted out the links to the previous reading notes at the end of the article. Few-shot semantic segmentation aims to learn to segment a new object class with only a few labeled samples, and most existing methods consider the setting of sampling base classes from the same domain as the new class (assuming the source domain and target domains are similar). However, in many applications it is not feasible to collect enough training data for meta-learning. This paper also extends small-sample semantic segmentation to a new task called cross-domain small-sample semantic segmentation (CD-FSS), which generalizes the meta-knowledge of domains with sufficient training labels to low-resource domains, and establishes CD-FSS. A new benchmark for FSS tasks.
Before starting to introduce CD-FSS, let's understand the concepts of cross-domain and small-sample learning in a broad sense (the articles later in this series will not be introduced in detail). Small sample learning can be divided into Zero-shot Learning (that is, to identify category samples that have not appeared in the training set) and One-Shot Learning/Few shot Learning (that is, in the training set, each category has one or several samples). Several related important concepts:
Domain: A domain D consists of a feature space X and a marginal probability distribution P(X) on the feature space, where X=x1,x2,…,xn, P(X) represents the distribution of X.
Task: Given a domain D={X, P(X)}, a task T consists of a label space Y and a conditional probability distribution P(Y|X)&