DeepLiDAR
DeepLiDAR copied to clipboard
JiaxiongQ
Deep Surface Normal Guided Depth Prediction for Outdoor Scene from Sparse LiDAR Data and Single Color Image (CVPR 2019)
Hi, I am using the scene_vis repo (linked [here](https://github.com/kujason/scene_vis)) to visualize the DeepLiDAR depth maps as 3D point-clouds and am encountering some odd visuals. I have attached two images below....
Hello, JiaxiongQ! when i read your code, i found that after you predict the surface normal vector, you perform some transformation operations on it. the code is: outputN = torch.zeros_like(normals2)...
Thank you for releasing the code. I have a question about generating dense depth. In 'depthCompletionNew_block' and 'depthCompletionNew_blockN' functions, the final output is generated by 'predict_normalE2'. However, depth map is...
hello when i try to reproduce the infrence result, i use the KITTI val_selection_cropped data, but the pred make me confused, when i print the pred.max(), i find all are...
Thanks for the excellent work. In your paper, I found that you train/test your model on NYU_v2 dataset and the delta metric are calculated, for delta(1.25^3), you reach 100%, I...
Hi! Thanks for your great work. How did you compute surface normal ground truth for carla data? From the ground truth dense depth map by local plane fitting?
你好： 首先很感谢能开源这样优秀的代码，但是有以下一些问题是我在复现你的训练过程中遇到的，希望你帮忙协助确认和解决，谢谢！ Q1:生成surface normal的代码，是从lidar depth和groundtruth depth生成的吧？且生成的surface normal都是3通道的？ Q2:进行第一步训练时，trainN.py的数据输入是rgb image，surface normal(lidar depth生成的)和surface normal(groundtruth depth生成的)。是用这三个数据作为输入吗？ Q3:我按照Q2的3个数据作为输入，发现你的trainN.py调用的trainLoader.py而不是trainLoaderN.py，是有错误吗？ Q4:我按照你的代码，使用trainN.py调用的trainLoader.py，读如的图片都是3通道的，为什么可以在代码中要把它reshape到1通道？是存在笔误吗？（如果我按照trainLoaderN.py，只有对其中一个surface normal作为多通道转单通道的处理，另外一个surface normal却没有，是什么原因呢）
Hello JiaxiongQ, Thank you for the wonderful work. I have some questions: 1. I want to know how to train on CARLA dataset? I didn't find the calibration files in...
I am unable to download the Carla dataset from the provided link: https://pan.baidu.com/s/1ayNWa7_9Ia2f6_lYzW8paA#list/path=%2F Do you have an alternate link?
您好！ 非常感谢您开源这个SOTA的代码。在训练过程中有几个问题如下： 1. 在我的任务里，需要尽量轻便的深度补全网络，因此打算只从头训练trainD,py，在dataloader里发现有三个fileppath。分别是filepathl、filepathd、filepathgt，前两个分别是KITTI depth数据集的data_depth_velodyne/train和data_depth_annotated/train文件夹，想知道第三个filepathgt在我的任务里应该是什么呢，或者是否可以做修改使得trainD.py进行独立的训练呢。 2. 训练完trainD.py后，有不同surface normal，仅在trainD.py上得到的model和结果进一步训练train.py来提高rmse或者其他metrics的可能性吗~？
