Towards weakly-supervised focus region detection via recurrent constraint network.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• State-of-the-art focus region detection (FRD) methods heavily depend on deep convolutional networks requiring expensive pixel-level annotations.
• Existing methods struggle with the trade-off between annotation cost and accuracy in FRD tasks.
• Bounding box annotations offer a more accessible labeling approach but lack precise boundary delineation for transition areas.

Purpose of the Study:

• To propose a novel FRD method that utilizes easily obtainable bounding box annotations.
• To achieve competitive accuracy comparable to fully supervised methods while significantly reducing annotation costs.
• To introduce a new dataset, FocusBox, to facilitate research in weakly supervised FRD.

Main Methods:

• Development of a recurrent constraint network (RCN) trained jointly with a fully convolutional network (FCN) using box-level supervision.
• Implementation of a dynamic training strategy involving iterative fine-tuning of FCN and RCN with generated pixel-level tags.
• Integration of a guided conditional random field to enhance the quality of generated pixel-level tags for improved boundary delineation.

Main Results:

• The proposed method achieves competitive accuracies in focus region detection using only bounding box annotations.
• Experimental results demonstrate comparable performance to fully supervised methods on existing datasets.
• The method achieves a faster processing speed compared to traditional fully supervised approaches.

Conclusions:

• Weakly supervised FRD using bounding box annotations is a viable and efficient alternative to pixel-level annotation methods.
• The proposed RCN and dynamic training strategy effectively address the challenge of boundary delineation with limited supervision.
• The FocusBox dataset provides a valuable resource for advancing research in weakly supervised FRD.