Abstract
We present a new object representation, called Dense RepPoints, which utilize a large number of points to describe the multi-grained object representation of both box level and pixel level. Techniques are proposed to efficiently process these dense points, which maintains near constant complexity with increasing point number. The Dense RepPoints is proved to represent and learn object segment well, by a novel distance transform sampling method combined with a set-to-set supervision. The novel distance transform sampling method combines the strength of contour and grid representation, which outperforms the counter-parts using contour or grid representations. Code is available at \url{this https URL}.
Abstract (translated)
URL
https://arxiv.org/abs/1912.11473