Goodpoint: unsupervised learning of key point detection and description

Performance Comparison of Feature Generation Algorithms for Mosaic Photoacoustic Microscopy

“…Unfortunately, the performance of these approaches is limited by the hardware performance, storage, and manual work [61]. New solutions are being investigated by applying deep learning algorithms for mosaic imaging, such as convolutional neural networks (CNNs) or GoodPoints [19]. Both methods extract feature localization and feature descriptors from a CNN model, which offers the same information but treats it differently.…”

“…Based on the CNN architecture backbone, A.V. Belikov [19] designed unsupervised learning processes for feature detectors and descriptors by following four stages: (1) warping two patches into the same size, (2) extracting feature positions and feature position using a two-headed CNN, (3) calculating descriptor loss as for all interpolated descriptors, (4) the feature location match with the descriptor was used as positive examples from detector training. The detector was trained with a loss function of position map L p by the summation of the feature's loss L f eatures and heatmap's loss L heatmaps .…”

“…Unfortunately, these traditional methods are limited by large databases or hand-tuning to find the models closest to the experiment. Thus, GoodPoint [19] is introduced as a balanced solution with deep learning-based feature generation to accelerate and learn without bulky algorithms to define the properties of the regions of interest. After generating the feature information, pairs of patches with overlapping areas were combined using the mosaic imaging process.…”