Fully Convolutional Network-Based Self-Supervised Learning for Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep learning models require large, annotated datasets, which are costly and time-consuming to acquire.
• Semantic segmentation, a key computer vision task, faces challenges due to data annotation limitations.

Purpose of the Study:

• To develop a novel self-supervised learning framework for representation learning from unlabeled data.
• To improve semantic segmentation performance when labeled data is scarce.

Main Methods:

• A self-supervised learning framework was developed by treating a jigsaw puzzle problem as a patch-wise classification task.
• A fully convolutional network was employed to solve the jigsaw puzzle with 25 patches.
• Learned features were transferred to the semantic segmentation task.

Main Results:

• Achieved a 5.8% performance improvement on the Cityscapes dataset compared to random initialization, using only 1/6 of the training images.
• Demonstrated the framework's applicability to different datasets and models, including PASCAL VOC2012.
• Achieved competitive results with state-of-the-art methods with reduced pretraining time costs.

Conclusions:

• Self-supervised learning via jigsaw puzzle solving is an effective method for representation learning in semantic segmentation.
• The proposed framework addresses the challenge of limited labeled data in computer vision tasks.
• This approach offers a computationally efficient alternative for achieving high performance in semantic segmentation.