샘플링된 깊이 추정 특징 맵을 이용한 의미론적 분할

초록

This paper proposes a multi-task learning network for semantic segmentation using feature maps from depth estimation networks. The feature map sampled from the depth estimation network was organized into multi-scale and handed over to the semantic segmentation network. This makes up a much lighter structure than the networks used for conventional semantic segmentation, while producing similar performance. Also in depth estimation, false depth results are reduced and noise was suppressed by learned semantic information. It achieved good performance for both tasks by simultaneously learning depth estimation and semantic segmentation, sharing information that cannot be retrieved, and being complementary to one another. Additionally, we showed that multi-task learning is possible for omnidirectional fisheye image dataset.

키워드