Augmented multiple instance regression for inferring object contours in bounding boxes.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Semantic segmentation models require extensive pixel-wise annotated contour data for training.
• Manual contour annotation is labor-intensive and costly, hindering model development.
• Existing methods often rely solely on bounding boxes or full contour data.

Purpose of the Study:

• To develop a cost-effective method for generating training data for semantic segmentation.
• To infer precise object contours from bounding boxes using limited contour data.
• To reduce the manual annotation effort in creating datasets for semantic segmentation.

Main Methods:

• Proposes Augmented Multiple Instance Regression (AMIR) to infer object contours.
• Treats bounding boxes as bags and contour hypotheses as instances.
• Leverages object contours to guide and regularize the Multiple Instance Regression (MIR) training process.

Main Results:

• AMIR accurately infers object contours within bounding boxes.
• The approach effectively utilizes a mix of contour and bounding box data.
• Evaluated on the Pascal VOC segmentation task, demonstrating promising performance.

Conclusions:

• AMIR offers an effective alternative to manual contour labeling for semantic segmentation.
• The method significantly reduces the cost and effort associated with data acquisition.
• Enables the creation of larger and more diverse training datasets for improved model performance.