Topic Title: Prior Knowledge Based Object Detection for IoT Devices


Technical Area:

Computer Vision
Deep Learning



In the past two decades, Internet is a crucial platform to connect people all around world. Besides more people, more things will access Internet in the future to construct the Internet of Things (IoT). In the meantime, artificial intelligence, especially deep learning, and high performance embedded computation chips have received great success in various domains. Enabling IoT devices with artificial intelligent characteristics helps to construct smart things. In contrast to the Cloud with a general AI ability, IoT devices focus more on specific and dedicated scenarios, such as smart home, intelligent transportation and so on. Traditional machine learning algorithms are able to incorporate prior knowledge to reduce the computation complexity of object detection and the error rate of object detection. This can refer to the best paper of CVPR2016 "Putting objects in perspective". Unfortunately, in the current stage, deep learning lacks such an ability to integrate prior knowledge to enable itself deployed on the IoT devices with a low inference computing complexity.



In most application scenarios, the backgrounds almost maintain constant, and the corresponding geometry knowledge could be obtained easily in advance. Therefore, the purpose of this topic is to integrate the prior knowledge into the object detection models, which aims to reduce the computation complexity of the inference process and eventually improve the accuracy of object detection.


Related Research Topic

1. Object detection with static backgrounds. This task is to quickly identify static backgrounds and dynamic target regions, and then to perform object detection and location within the interested target regions.

2. Object detection in perspective view. From the perspective view, objects look small from a far distance and become large after stepping in a short distance. Thus, our goal is to incorporate the geometry knowledge into the detection process to search the target objects, instead of sliding window based brute force searching.