SRN: Side-Output Residual Network for Object Reflection Symmetry Detection and Beyond

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Object reflection symmetry detection is crucial for image understanding.
• Existing datasets lack the complexity of real-world scenes, hindering progress.

Purpose of the Study:

• To establish a robust benchmark for reflection symmetry detection in natural images.
• To propose an effective deep learning model for accurate symmetry detection.

Main Methods:

• Introduction of the Sym-PASCAL benchmark, featuring diverse objects, occlusions, and complex backgrounds.
• Development of a side-output residual network (SRN) utilizing residual units (RUs) for error fitting across network stages.
• Extension to a multitask SRN (MT-SRN) for joint symmetry and edge detection.

Main Results:

• Sym-PASCAL proves to be a challenging benchmark, reflecting real-world image complexities.
• The proposed SRN achieves state-of-the-art performance in reflection symmetry detection.
• MT-SRN demonstrates effective joint prediction of symmetry and edge masks without performance degradation.

Conclusions:

• The Sym-PASCAL benchmark provides a valuable resource for advancing symmetry detection research.
• The SRN architecture offers a powerful and generalizable approach for image-to-mask learning tasks.
• Multitask learning enhances model versatility and efficiency in computer vision applications.