Segmenting Objects in Day and Night: Edge-Conditioned CNN for Thermal Image Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Visible spectrum imaging faces limitations in adverse environmental conditions.
• Thermal infrared cameras offer advantages like night vision and insensitivity to illumination.
• Semantic segmentation in thermal imagery is crucial for day and night applications.

Purpose of the Study:

• To propose a novel network architecture for enhanced thermal image semantic segmentation.
• To improve segmentation quality by adaptively incorporating edge prior knowledge.
• To introduce a comprehensive benchmark dataset for evaluating thermal image segmentation methods.

Main Methods:

• Developed the edge-conditioned convolutional neural network (EC-CNN).
• Designed a gated featurewise transform layer to integrate edge information.
• Introduced the Segmenting Objects in Day And night (SODA) dataset with 7168 annotated thermal images.
• Conducted end-to-end training of the EC-CNN model.

Main Results:

• EC-CNN generated high-quality semantic segmentation results with edge guidance.
• The proposed EC-CNN demonstrated superior performance compared to state-of-the-art methods.
• Experiments on the SODA dataset validated the effectiveness of EC-CNN.

Conclusions:

• EC-CNN is an effective architecture for thermal image semantic segmentation.
• The integration of edge priors significantly enhances segmentation quality.
• The SODA dataset provides a valuable resource for advancing research in this field.