Open yunxiaozimeng opened 3 years ago
@yunxiaozimeng The dataset for training only requires rectified stereo-pairs of images. Getting real rectified stereo pairs may be difficult to set-up, so instead you could get some from a simulated environment like airsim (if you don't care what the subject is). No annotation is required as it is a self-supervised framework! All you have to do is to organise the filenames similar to the methods in utils/filenames to left-right pairs.
@yunxiaozimeng用于训练的数据集仅需要校正的图像立体对。要获得真实的整流立体声对可能很难设置,因此您可以从模拟环境(如airsim)中获得一些(如果您不在乎对象是什么)。由于它是一个自我监督的框架,因此无需注释!您要做的就是将类似于utils / filenames中的方法的文件名左右组合。
thank you very much,your answer is helpful !
I CAN'T use KITTI datasets because my terrible download speed,so I want to create my own dataset but I don't know the format of dataset in this project, who can tell me which files of datasets i need and how can i annotate images? THANK YOU!!