A Survey on Object Detection Performance with Different Data Distributions

Abstract

Detecting objects in a dynamic scene is a critical step for robotic navigation. A mobile robot may need to slow down in presence of children, elderly or dense crowds. A robot’s movement needs to be precise and socially adjustable especially in a hospital setting. Identifying key objects in a scene can provide important contextual awareness to a robot. Traditional approaches used handcrafted features along with object proposals to detect objects in images. Recently, object detection has made tremendous progress over the past few years thanks to deep learning and convolutional neural networks. Networks such as SSD, YOLO, and Faster R-CNN have made significant improvements over traditional techniques while maintaining real-time inference speed. However, current existing datasets used for benchmarking these models tend to contain mainly outdoor images using a high-quality camera setup that is usually different from a robotic vision setting where a robot moves around in a dynamic environment resulting in sensor noise, motion blur, and change in data distribution. In this work, we introduce our custom dataset collected in a realistic hospital environment consisting of distinct objects such as hospital beds, tables, and wheelchairs. We also use state-of-art object detectors to showcase the current performance and gaps in a robotic vision setting using our custom CHART dataset and other public datasets.