상세정보
학술대회
Depth Attention Net
- 발행일
- 201910
- 출처
- International Conference on Information and Communication Technology Convergence (ICTC) 2019, pp.1110-1112
- 협약과제
- 19PS2300, 세라믹산업 제조혁신을 위한 클라우드 기반 빅데이터 플랫폼 개발, 지수영
- 초록
- Depth estimation has been achieved much attention in recent five years for visual SLAM, AR, VR, autonomous vehicle, 3D object understanding and so on. Similar to stereo matching analysis, stereo image based depth estimation methods using deep learning surprisingly obtain great improvement in performance but still cannot understand why learning depth can estimate depth in a new image and also require accuracy improvement. Therefore, we apply attention mechanism into depth learning. Attention mechanism achieve great improvement in various applications such as visual identification of objects, speech recognition, reasoning, image captioning, summarization, segmentation, machine translation(or NLP) and image classification etc. but yet depth estimation. This is because depth estimation is considered as geometric area. We apply attention mechanism in the PSMNet [1] method. Attention mechanism usually apply to channel attention. In our experiments, we apply our method into two benchmark dataset: KITTI [2] and SceneFlow [3]. In our experiments, we found the attention net can improve the quality of depth.
- KSP 제안 키워드
- 3D object understanding, Attention mechanism, Autonomous vehicle, Benchmark datasets, Depth estimation, Estimation method, Image captioning, Image classification, Machine Translation(MT), Matching analysis, Visual SLAM