電気通信システムと管理

Custom datasets from 0.-A brief discussion on how to acquire and prepare data for training deep learning neural networks. Whether for sorting, recognizing objects or targeting.

Abstract

André Costa

A quick chat about creating custom datasets from absolute 0.  The true path between the stones. (Focusing on data acquisition and preparation. Without going into details of neural network training). A summary of what I wish someone had talked to me when I was starting ... It will be addressed from the collection of images. Criteria for image acquisition, such as ambient light control, dark areas in the image. Existence of hotspots. The importance of highlighting the points of the scene with really relevant data. Possible improvements in the quality of images collected using optical filters to alleviate major problems. Such as eliminating light reflections, enhancing or eliminating colors. What are the main filters that should always be in our toolbox? Imaging from special cameras such as infrared cameras, multispectral cameras and thermal cameras. The least I need to know about DICOMS. Do I need special and expensive software to deal with dicoms? Quantities of images to train a network. - "Why do I need two thousand images if I trained my network with only one image?". Data Standardization (Microsoft COCO, Pascal VOC). Main tools (free and paid) to annotate images for object recognition and segmentation. Ok. I prepared my dataset. And now? How do I do to train? Some frameworks for network training. Image classification classification, recognition and segmentation of objects in images. Gpus. What is this? Do I need this to train? And after training? Is it mandatory to use gpus to run my project? Now I know how to train a network! I already know everything! Am I ready for the market? Only not ....