Practical Analysis for Improving Performance from One-stage Object Detectors

Abstract

Since the advent of the CNN, the performance of object detectors has been greatly improved. In addition, with the one-stage object detector, the detection algorithm has been lightened and improved to a level that can be applied to real-time applications. However, the research directions for one-stage object detectors focus on obtaining high performance on a benchmark dataset, and there is less consideration of how to improve performance with real-world data. In this paper, we check how methods popularly used to enhance performance from neural networks respond to datasets similar to real-world data. Also, we experimentally confirm that a training configuration setup that considers the target dataset can be more effective than complex training strategies like knowledge distillation. Although this analysis is somewhat heuristic, we expect it will provide meaningful insights for researchers and developers who want to apply an object detector to actual applications.