object detection , image classification , neural networks , YOLO , optimizers

With the increased accessibility to neural network frameworks and computation clouds, a wide range of competition websites offer real tasks for the neural network community to join. In this paper, we discuss a problem of object detection and classification. Based on our experience, we describe a general pipeline and necessary steps that may help researchers willing to participate in such competition. We partition the problem into two separate tasks. Firstly, we present state of the art for relevant neural networks concerning accuracy and computation time trade-off. Further, we create a survey of major techniques that leads to accuracy improvement. Namely, we recall image augmentation techniques, demonstrate the impact of various optimizers, and discuss ensemble techniques. The pipeline and techniques reflect our experience with a competition, in which we were able to reach a highly competitive solution and ended in fourth place. The uniqueness of our solution is that we used only free Google Colab computation service and still overperformed many more computation extensive approaches.