Large Scale Shadow Annotation and Detection Using Lazy Annotation and Stacked CNNs.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Shadow detection is crucial for image analysis but hindered by the scarcity of annotated data.
• Existing methods struggle with broad image domains due to the labor-intensive annotation process.

Purpose of the Study:

• To introduce an efficient annotation method called "lazy annotation" for shadow detection.
• To develop a robust shadow detection model capable of generalizing across diverse image datasets.

Main Methods:

• Proposed "lazy annotation" method, requiring annotators to mark only key shadow and non-shadow areas, resulting in noisy labels.
• Developed a joint learning framework to address label noise by simultaneously training a shadow region classifier and recovering corrupted labels using Least Squares Support Vector Machines (LS-SVM).
• Introduced a stacked Convolutional Neural Network (CNN) architecture trained on a large, diverse dataset, incorporating image-level semantic information for patch refinement.

Main Results:

• Classifiers trained with recovered labels achieved performance comparable to those trained on fully annotated data.
• A new, large-scale dataset was created, 20 times larger than existing ones, featuring diverse scenes and image types.
• The proposed stacked CNN pipeline, trained with recovered labels, achieved state-of-the-art performance and demonstrated strong generalization capabilities on cross-dataset tasks.

Conclusions:

• "Lazy annotation" is an effective strategy for creating large-scale shadow detection datasets with reduced annotation effort.
• The proposed stacked CNN architecture and dataset enable highly accurate and generalizable shadow detection.
• The research significantly advances the field of shadow detection by addressing data scarcity and improving model robustness.