A Comparative Study for Fisheye Image Classification: SVM or DNN

The comparison between the feature-based method and the learning-based method is conducted in the training time, the accuracy and the generalization capacity, to address the optimisation for the multi-style fisheye imagery classification. We construct an srd-SIFT descriptor based SVM classifier to present the feature-based method for describing the influence of the dataset scale and the visual word scale on the classifier. The SVM classifier achieves 15.98% accuracy on the test set after 162 h training, with the condition that includes 800 images per class in 12 classes and 1500 visual words. For the learning-based method, we propose to expand training samples’ style variety, via style transformation, to facilitate the contemporary architecture retraining. Following this approach, we retrain the ResNet-50 by an artificial multi-style fisheye image dataset without complementing new training labels. The performance of the obtained ResNet classifier is evaluated on 6000 images collected in the real-world. The result shows that the retrained classifier has great generalization capacity and reaches 97.19% top-3 accuracy.