Abstract: Much fruitful study has been conducted on object detection which is one of the
fundamental problems in deep learning. Self-service freezer is an important application of artificial
intelligence in the retail industry. Object detection methods are used to detect goods in pictures
captured by cameras inside the freezer, and tasks such as commodity classification follow suit. Due to
the limitation of hardware, currently we only apply fast while less accurate models in practical
application, of which the detection accuracy is much worse, for small objects. In an attempt to explore
the features of data collected in self-service freezers such as single background and small range of
object, a coarse-to-fine two-stage method called FocusNet was proposed to tackle the problem of
object detection under this special condition, which was based on the previous main stream one-stage
detection method. The experimental results show that FocusNet outperforms the previous method by
about 8.3% and 3.5% in small object detection and overall detection, respectively.

Key words: object detection, small object detection, coarse-to-fine, self-service freezer, deep learning