FishSegSSL: A Semi-Supervised Semantic Segmentation Framework for Fish-Eye Images

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Large field-of-view (FoV) fish-eye cameras offer advantages in applications like autonomous driving.
• Deep learning for computer vision typically relies on large labeled datasets, which are scarce for fish-eye imagery.
• Semi-supervised learning presents a viable approach to address data limitations in fish-eye image analysis.

Purpose of the Study:

• To explore and benchmark existing semi-supervised learning methods for fish-eye image segmentation.
• To introduce FishSegSSL, a novel framework designed for semi-supervised semantic segmentation of fish-eye images.
• To evaluate the effectiveness of FishSegSSL on a real-world dataset from vehicle-mounted cameras.

Main Methods:

• Benchmarking two established semi-supervised learning techniques in the context of fish-eye images.
• Developing the FishSegSSL framework incorporating pseudo-label filtering, dynamic confidence thresholding, and robust augmentation.
• Utilizing the WoodScape dataset, which contains images from vehicle-mounted fish-eye cameras.

Main Results:

• FishSegSSL achieved performance improvements of up to 10.49% over fully supervised methods with equivalent labeled data.
• The proposed method enhanced existing image segmentation techniques by 2.34%.
• This research represents the first investigation into semi-supervised semantic segmentation specifically for fish-eye images.

Conclusions:

• Semi-supervised learning is effective for fish-eye image segmentation, overcoming data scarcity challenges.
• The novel FishSegSSL framework demonstrates significant performance gains and robustness.
• Further ablation studies and sensitivity analyses confirm the efficacy of the individual components within FishSegSSL.