This package implements the C3D loss (continuous 3D loss) in the paper Monocular Depth Prediction Through Continuous 3D Loss. The paper is accepted by IROS 2020.

Video Demo:

This repo contains the python package for computing Continuous 3D Loss. For complete networks conducting monocular depth prediction trained with this loss, please check this DORN implementation.

Implementation of the C3D loss in BTS and Monodepth2 network will be released soon.

The interface for using the C3D loss is the C3DLoss class in c3d_loss.py. You can import the class to your script to augment your own depth prediction project with C3D loss. The implementation of its operations are in cvo_funcs.py and cvo_ops folder. The class also needs some other util classes to run properly, which you can find in utils folder.

• Python 3.6
• Pytorch 1.2
• PIL

This is the environment I used to work properly. Higher versions are likely to be fine, but I did encountered problem of divergence in no_grad mode when working on another environment with high version of Pytorch. The exact problem is not located yet.

To use the C3D loss, you need to install this library.

$ python setup.py install

1. You will need to create an instance for these classes:

• C3DLoss (for calculating the proposed loss),
• CamProj (for generating camera intrinsic and extrinsics information needed by C3DLoss), and
• DataReaderKITTI (enabling easier retrieval of camera intrinsic and extrinsics information from KITTI dataset).

2. C3DLoss can be simply initialized as c3d_loss = C3DLoss(). Optionally you can create a config text file to customize the parameters. See c3d_config_example.txt for an example. C3DLoss.parse_opts(f_input=$path_to_config_file$) set the parameters specified in the config file.
3. CamProj and DataReaderKITTI objects should be initialized in the dataset-handling part of your code. DataReaderKITTI takes the root path to your local KITTI dataset as initialization argument. CamProj takes the DataReaderKITTI object as initialization argument.
4. Camera intrinsic and extrinsics information are encapsuled in the CamInfo class, which is output by a CamProj object. CamInfo object is needed in C3DLoss forward calculation, together with RGB image, depth prediction, ground truth, and validity masks.

Please check here as an example of using the C3D loss in an existing network (DORN).

This repo is still under constructions. The implementation is subject to change. More comments and instructions would come later. Your comments are welcomed.