The accuracy of visual object detection that estimates
locations and classes of target objects in input images
has been drastically improved by rapidly advancing
technology
about deep convolutional neural networks (CNNs). The
evaluation of existing methods based on CNNs is usually
conducted
using major datasets such as MS-COCO, PASCAL-VOC, etc:
these datasets include several sizes of target objects. The
accuracy of detection larger objects has become excellent
by
recent methods but it has been still difficult even for
recent
CNNs to detect small object accurately. To solve this
problem,
this paper investigates how to improve the accuracy of
small
object detection with CNNs. For the investigation, two
kinds of
datasets composed only of small target objects were
created:
the bird dataset including only flying objects in the sky
and
the SAVMAP dataset having only mammals on the savannah.
Experimental results using the datasets showed that input
size,
depth of CNN layers, and surrounding context of target
objects
were important factors for small object detection.
Experimental
results showed that, EfficientDet-D0 achieved an accuracy
of
0.6585 for the bird dataset and 0.6501 for the SAVMAP
dataset.