人工智能培训

搜索

人工智能论文:基于非对称注释的实时联合语义分割和深度估计(Real-Time Joint Semantic Segmentation and Depth Est

[复制链接]
rrrrrrr 发表于 2018-9-14 08:55:43 | 显示全部楼层 |阅读模式
rrrrrrr 2018-9-14 08:55:43 115 0 显示全部楼层
人工智能论文:基于非对称注释的实时联合语义分割和深度估计(Real-Time Joint Semantic Segmentation and Depth Estimation Using  Asymmetric Annotations)在机器人技术中作为感知信息提取器部署深度学习模型可能是一项艰巨的任务,即使使用通用GPU卡也是如此。在这里,我们解决了三个最突出的障碍,即i)适应单个模型同时执行多个任务(在这项工作中,我们考虑深度估计和语义分割对于获取场景的几何和语义理解至关重要),而ii)实时进行,以及iii)使用每种模态具有不均匀数量的注释的不对称数据集。前两个问题,我们采用最近提出的实时语义分段网络,进行少量更改以进一步减少浮点运算的数量。为了解决第三个问题,我们采用了一种基于硬知识蒸馏的简单解决方案,假设有一个强大的“教师”网络。最后,我们展示了我们的系统如何能够轻松扩展以同时处理更多任务和更多数据集。定量地,我们在NYUDv2上使用单个模型实现42%的平均值,0.56m RMSE(lin)和0.20 RMSE(log) -40,87%表示iou,3.45m RMSE(lin)和0.18 RMSE(log)在KITTI-6上用于分割,KITTI用​​于深度估计,前向通过成本仅为17ms,在1200x350输入上为6.45 GFLOP。所有这些结果要么等同于(或者更好)当前最先进的方法,这是通过较大和较慢的模型分别解决每个任务来实现的。
Deployment of deep learning models in robotics as sensory informationextractors can be a daunting task to handle, even using generic GPU cards.Here, we address three of its most prominent hurdles, namely, i) the adaptationof a single model to perform multiple tasks at once(in this work, we considerdepth estimation and semantic segmentation crucial for acquiring geometric andsemantic understanding of the scene), while ii) doing it in real-time, and iii)using asymmetric datasets with uneven numbers of annotations per each modality.To overcome thefirst two issues, we adapt a recently proposed real-timesemantic segmentation network, making few changes to further reduce the numberof floating point operations.To approach the third issue, we embrace a simplesolution based on hard knowledge distillation under the assumption of havingaccess to a powerful `teacher' network.Finally, we showcase how our system canbe easily extended to handle more tasks, and more datasets, all at once.Quantitatively, we achieve 42% mean iou, 0.56m RMSE (lin) and 0.20 RMSE (log)with a single model on NYUDv2-40, 87% mean iou, 3.45m RMSE (lin) and 0.18 RMSE(log) on KITTI-6 for segmentation and KITTI for depth estimation, with oneforward pass costing just 17ms and 6.45 GFLOPs on 1200x350 inputs.All theseresults are either equivalent to (or better than) current state-of-the-artapproaches, which were achieved with larger and slower models solving each taskseparately.人工智能论文:基于非对称注释的实时联合语义分割和深度估计(Real-Time Joint Semantic Segmentation and Depth Estimation Using  Asymmetric Annotations) VBbFybgdHaOG3ORl.jpg
URL地址:https://arxiv.org/abs/1809.04766     ----pdf下载地址:https://arxiv.org/pdf/1809.04766    ----人工智能论文:基于非对称注释的实时联合语义分割和深度估计(Real-Time Joint Semantic Segmentation and Depth Estimation Using  Asymmetric Annotations)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

rrrrrrr当前离线
新手上路

查看:115 | 回复:0

快速回复 返回顶部 返回列表