研究目的
Investigating the possibility of learning to directly output a vectorial semantic labeling of the image by casting a mapping problem as a polygon prediction task.
研究成果
The proposed PolyCNN network succeeds in learning to predict vertices of 4-sided polygons corresponding to the objects of interest, yielding better results than a 2-step process involving a U-Net followed by vectorization. Future work includes experimenting with different feature extractors and learning n-sided polygons with variable n.
研究不足
The study is restricted to learning quadrilaterals or 4-sided polygons. The proposed architecture is relatively big with more than 2.7×10e6 weights, indicating potential areas for optimization in model size and complexity.
1：Experimental Design and Method Selection:
The study proposes a deep learning approach, PolyCNN, which predicts vertices of the polygons outlining objects of interest. The network architecture includes a feature extractor, encoder, and decoder.
2：Sample Selection and Data Sources:
The Solar photovoltaic array location dataset (PV dataset) is used, containing geospatial coordinates and border vertices for over 19000 solar panels across 601 high-resolution aerial orthorectified images.
3：List of Experimental Equipment and Materials:
The study utilizes a pre-trained InceptionV4 network for feature extraction and a custom decoder for polygon prediction.
4：Experimental Procedures and Operational Workflow:
Image patches are extracted around each polygon, data augmentation is performed, and the network is trained using an Adam optimizer.
5：Data Analysis Methods:
The performance is evaluated using Intersection over Union (IoU) and a polygon accuracy measure based on L2 distance between vertices.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
