Abstract
DISC introduces a progressive coarse-to-fine framework for salient object detection. By utilizing two stacked convolutional neural networks (CNNs)—one for global context and another for local refinement—the model generates pixel-accurate saliency maps. The integration of superpixel-based pooling layers, specifically Intra-superpixel smoothing (ISS) and Inter-superpixel voting (ISV), ensures spatial consistency. This data-driven approach allows the model to outperform traditional hand-crafted priors and achieve state-of-the-art performance on major benchmarks.
Introduction
In computer vision, "saliency" refers to the identification of the most informative or attention-grabbing regions in an image. Historically, researchers relied on mathematical priors, such as center-bias or high local contrast. However, these hand-crafted assumptions are often brittle; local contrast methods may only highlight boundaries, while global methods can fail in cluttered environments. DISC addresses these limitations by moving toward a purely data-driven, end-to-end learning framework capable of handling complex backgrounds and preserving fine object details.
Methodology: Coarse-to-Fine Progressive Learning
The core innovation of DISC is its two-stage architecture designed to mimic human visual perception—glancing at a scene to find regions of interest before focusing to resolve details.
Coarse-Level and Fine-Level Networks
The process begins with the Coarse-Level CNN, which analyzes the entire image to determine the "where" and "what" of salient objects. To ensure the output respects object boundaries, the authors integrate Superpixel-based Local Context Information (SLCI) via two novel pooling layers: Intra-superpixel smoothing (ISS), which averages saliency scores within regions for consistency, and Inter-superpixel voting (ISV), which refines scores based on regional similarity.
Subsequently, the Fine-Level CNN focuses on local patches guided by the coarse map. It processes the original image patch, the corresponding coarse saliency patch, and a "spatial regularization" map to produce a high-resolution, pixel-accurate final map.
Technical Innovations
Beyond the progressive structure, DISC introduces several key technical refinements:
- Hinge Loss Integration: Instead of standard classification, the authors embed a Support Vector Machine (SVM)-like hinge loss directly into the CNN, improving visual quality and accuracy.
- Non-parametric Spatial Regularization: Rather than hard-coding a center-bias prior, the network learns spatial distributions implicitly by using an average of all ground-truth maps from the training set as an additional input channel.
Experimental Results
The model was evaluated on major benchmarks including MSRA10K, ECSSD, and PASCAL1500. DISC demonstrated exceptional robustness, performing well across diverse datasets without fine-tuning, suggesting it learns fundamental features of saliency rather than just memorizing specific datasets.
Quantitatively, on the MSRA10K dataset, DISC achieved a precision of 97.3% and a 15.3% improvement in recall over previous state-of-the-art methods. As shown in the comparison above, DISC (column c) consistently produces maps that are closer to the ground truth (column b) than traditional methods based on local or global contrast (columns d-e). Furthermore, the system is efficient, processing maps in approximately 75ms on a single GPU. The authors also demonstrated task-oriented flexibility; by fine-tuning on specific categories like "butterflies," the model could be adapted from generic saliency to specific object detection.
Conclusion
DISC successfully bridges traditional computer vision techniques like superpixels with modern deep learning. While the fine-level map depends on the quality of the coarse-level output, the framework represents a significant step forward in data-driven saliency detection. Future work may involve incorporating higher-level semantic knowledge and extending the model to video saliency, where temporal consistency becomes critical.
Summary updated