论文笔记_S2D.61_2019-CVPR-DeepLiDAR:基于稀疏激光雷达数据和单张彩色图像的户外场景的表面法线引导的深度预测

悠悠 2023-01-01 03:55 141阅读 0赞

# 基本情况 #

*  题目：DeepLiDAR: Deep Surface Normal Guided Depth Prediction for OutdoorScene from Sparse LiDAR Data and Single Color Image
 *  论文地址：https://arxiv.org/abs/1812.00488
 *  来源：中国电子科技大学，旷视科技
 *  原文作者：Jiaxiong Qiu
 *  出处：Qiu J, Cui Z, Zhang Y, et al. Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image\[C\]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019: 3313-3322.

# 介绍 #

在本文中，提出了一种深度学习架构，它可从一张彩色图像和一个稀疏深度图生成室外场景之下的精确的稠密深度。受室内深度补全的启发，网络把表面法线估计作为中间表示，以产生稠密深度，并可以端到端训练。该架构采用改进的**编解码结构**，有效地融合了密集的彩色图像和稀疏的激光雷达深度。为了解决室外特定的挑战，该模型还预测一个**置信度掩膜**，以处理由于**遮挡**而造成的**前景边界附近**混合的激光雷达信号，并整合来自彩色图像的估量和带有已学习的**注意力图**的曲面法线，以提升深度的精度，尤其是远距离区域。

大量的实验表明，我们的模型在 KITTI 深度补全基准上的SOTA。同时，消融实验表明每个模型组件对最终结果的有效作用。综合分析证明该模型可以较好地推广到带有较高稀疏性的输入或者来自室内场景的输入。

下面是论文具体框架结构以及实验结果：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 1][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 2][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 3][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 4][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 5][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 6][]

//TODO

# 参考： #

[【深度估计】旷视科技|DeepLiDAR从一张彩色图像和一个稀疏深度图像生成室外场景之下的精确的稠密深度图][DeepLiDAR]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70]: /images/20221120/5a722111c29a42939285b0473dda7875.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20201231122303577.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20201231122326965.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/2020123112235286.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20201231122415515.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 5]: https://img-blog.csdnimg.cn/20201231122439945.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ_size_16_color_FFFFFF_t_70 6]: https://img-blog.csdnimg.cn/20201231122542600.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3NoeWpoeXAxMQ==,size_16,color_FFFFFF,t_70
[DeepLiDAR]: https://cloud.tencent.com/developer/article/1653134