研究目的
To extend the scale of the dataset for training deep neural networks on fisheye images by using parameterized synthetic images to boost sample diversity and avoid scale limitations.
研究成果
The proposed method of synthesizing fisheye images from labeled perspective images is effective for training deep neural networks, achieving high accuracy in classification tasks. It provides a scalable solution to overcome the limitations of collecting and labeling real fisheye images. The synthetic dataset is the first of its kind and is publicly available for further research.
研究不足
The method relies on existing labeled datasets, which may not cover all possible real-world variations. The projection model used is equidistant, which might not accurately represent all types of fisheye lenses. The synthetic images may lack some real-world noise and distortions.
1：Experimental Design and Method Selection:
The paper uses a parameterized transformation method based on equidistant projection geometry to synthesize fisheye images from existing labeled perspective images. It involves applying controllable projection processes to simulate different viewing angles and distances.
2：Sample Selection and Data Sources:
The ILSVRC 2012 training dataset with
3：2 million labeled images in 1000 categories is used as the raw image dataset. For validation, 12 classes of indoor objects are selected and transformed. List of Experimental Equipment and Materials:
A Dell Precision T5500 workstation with dual Intel Xeon Processor E5645 and MSI Geforce GTX 1080 Graphic card is used for computation. The Deep Learning Library Caffe is employed for implementation. For real image capture, a Fujifilm Fujinon F-FE185C057HA-1 lens and Point Grey Grasshopper 3 camera are used.
4：Experimental Procedures and Operational Workflow:
Raw perspective images are scaled to 512x512 pixels. Equidistant projection is applied with center offsets to generate synthetic fisheye images. The synthetic dataset is divided into training, validation, and test sets. A pre-trained Alexnet neural network is fine-tuned on this dataset, and its performance is evaluated on real fisheye images captured with the specified equipment.
5：Data Analysis Methods:
The accuracy of the neural network is assessed using top-1 and top-3 prediction reliability metrics on the test dataset and captured images.
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
