maximise mutual information betweenthe class assignments of each pair
It's the core idea.
5 Matching Annotations
- Jun 2023
-
www.semanticscholar.org www.semanticscholar.org
-
-
directly outputs semantic labels
Unlike STEGO, IIC does not need an extra cluster step.
-
- Apr 2023
-
www.semanticscholar.org www.semanticscholar.org
-
learning dense vi-sual representations from unlabeled scene-centric data
The main challenging that the paper aims to conquer.
-
Indeed, in the presenceof complex scene images, the random cropping operationused as image transformation loses its semantic-preservingproperty, as a single image can yield two crops bearing an-tipodean semantic content [31,35–37]. Along the same line,it’s not clear how to relate sub-regions of the image fromone crop to the other, which is necessary to derive a local-ized supervisory signal.
Open Issues: 1. Random cropping hurts the semantic consistency and leads to ambiguous semantic content when the original image contains complex scene . 2. 2 image crops from one image hold little localized supervisiory signal.
How to solve them?
-
More importantly, theclustering algorithm conjointly operates on the features ofboth views, thereby elegantly bypassing the issue of contentnot represented in both views and the ambiguous matchingof objects from one crop to the other.
What's that issues?
-