Abstract
The Guided Filter Network (GFN) is a novel framework designed to automatically generate high-quality, pixel-level image labels, drastically reducing the need for manual annotation in semantic segmentation. By transferring segmentation knowledge from source domains and using an iterative "self-improvement" loop, GFN produces fine-grained masks that rival human-labeled ground truth.
Introduction
Deep Convolutional Neural Networks (DCNNs) have revolutionized semantic segmentation, but they are data-hungry. To perform well, they require massive datasets with precise, pixel-level labels. Creating these labels is a "dreadfully labor-intensive" process that often results in human inconsistency. While "weakly-supervised" methods exist—using only bounding boxes, scribbles, or image-level tags—they often struggle to capture precise object boundaries. The Guided Filter Network (GFN) addresses this bottleneck by generating accurate "pseudo-labels" to train high-performance segmentation models without manual intervention.
Methodology
The GFN operates in two main stages: a proxy stage (learning from a source) and an iterative learning stage (refining in the target). The architecture integrates three key modules to ensure high-quality mask generation.
Core Modules
- Boundary Guided (BG) Module: This module highlights object edges using a boundary extractor. It calculates the difference between max-pooling and min-pooling operations to identify sharp transitions in feature maps.
- High-level Guided Low-level (HGL) Module: To mitigate noise in low-level features, this module uses high-level semantic features as a gate, ensuring the network focuses on relevant structural details.
- Guided Filter (GF) Module: An edge-preserving component that uses the original image as a guide to sharpen the final mask, aligning it perfectly with object boundaries.
Training Strategies
To initialize the system without target labels, GFN employs three pretext tasks:
- Transfer Learning: Leveraging knowledge from semantically similar source domains.
- Simple-to-Complex: Training on images with simple backgrounds (using Otsu’s method) before tackling complex scenes.
- Class Activation Maps (CAMs): Refining image-level labels into initial regional masks.
The process concludes with Iterative Learning, where coarse masks are treated as pseudo-labels to progressively re-train and refine the network.
Results
GFN was evaluated across six diverse datasets, including PASCAL VOC 2012, FGVC Aircraft, and CUB-200-2011.
- Quantitative Performance: On PASCAL VOC 2012, GFN achieved a mean Intersection over Union (mIoU) of 62.3%, surpassing many state-of-the-art weakly-supervised models. On specialized datasets like FGVC Aircraft, it reached an mIoU of 91.5%.
- Qualitative Quality: Visual results show that the predicted masks are remarkably sharp, capturing intricate structures like bird beaks and tree branches that typically challenge automated systems.
- Scalability: Performance improved significantly as more unlabelled "simple" images were added, demonstrating the framework's ability to scale with data.
Conclusion
The Guided Filter Network provides a robust solution for automatic pixel-level labeling. By combining boundary guidance with semantic filtering and iterative self-correction, it bridges the gap between weakly-supervised methods and manual annotation. Future work may explore integration with Transformer-based backbones and more sophisticated reliability metrics to further mitigate semantic drift during iterative learning.
Summary updated