search

HKUST-Aerial-Robotics / MVDepthNet

This repository provides PyTorch implementation for 3DV 2018 paper "MVDepthNet: real-time multiview depth estimation neural network"
GNU General Public License v3.0
309 stars 72 forks source link

how could I train my model or using other image size? #1

Closed ahccom closed 5 years ago

ahccom commented 5 years ago

Good job! But my image size is (372,240),how could I use the network to fitting it? Or how could I train the network on my dataset? Looking forward to your reply, thanks.

WANG-KX commented 5 years ago

Thanks for your interest in this work! The input image size can be changed as long as the camera intrinsic parameters are changed accordingly. You can take a look at the example code. The core part is how to construct the coat function according to the paper. If I have time, I will provide an OpenCV based code that is easier to understand.

WANG-KX commented 5 years ago

Training the network is quite straightforward, no fancy tricks. All details are provided in the paper. As far as I think, the main difficulty is to collect the dataset. You can download it from DeMoN. Also, it takes days to train.

chenchr commented 5 years ago

Hello. Thanks for the code. Did you undistort the image of rgbd-slam dataset ? I am doing a learning based mapping project now. A question is how to adjust the image distortion. To pre-undistort before training or consider the distortion parameter during warpping ? Besides, as the size of image in different dataset differs, did you resize the image and intrinsic to a canonical size ? Or manually switch between different dataset when training ? Thanks !

WANG-KX commented 5 years ago

Dear, Thanks for your interest :) I assume the images from TUM RGB-D DATASET are undistorted. During the training and testing, images are resized to the same resolution (320×256). The intrinsic parameters are changed accordingly. Regards, Kaixuan