Deep Neural Networks are powerful at solving classification problems in computer vision.
However, learning classifiers with these models requires a large amount of labeled training
data, and recent approaches have struggled to adapt to new classes in a data-efficient manner.
There is interest in quickly learning new concepts from limited data using one-shot learning
methods [21, 37]. One-shot image classification is the problem of classifying images given
only a single training example for each category [22, 39].
We propose to undertake One-Shot Semantic Image Segmentation. Our goal is to predict
a pixel-level segmentation mask for a semantic class (like horse, bus, etc.) given only a single
image and its corresponding pixel-level annotation. We refer to the image-label pair for the
new class as the support set here, but more generally for k-shot learning, support set refers
to the k images and labels.
...
На что обратить внимание там ниже?