semonemo / STAD

STAD: Stable Video Depth Estimation (ICIP 2021)

Getting Started

This code has been developed under Anaconda(Python 3.6), Pytorch 1.6.0, Torchvision 0.7.0 and CUDA 10.1.

1. Please install following environments:

   # create a new environment if needed
   conda create --name stad
   conda activate stad
   
   # install the the dependencies 
   pip install -r requirements.txt
   

2. Download KITTI data

   (1) Download raw data

   • Option1. Follow the instruction from BTS github
   • Option2. Download the raw dataset from KITTI dataset page

   (2) Download ground truth data

   • Download official ground truth from KITTI dataset page

3. Download pretrained weights

   • From google drive

Train on KITTI

In the code folder, run

sh local_train_kitti.sh

You need to change dataset_path, exp_name and loss_type correctly.

Please set parameters as follows:

Models	exp_name	loss_type
Neural-RGBD	ver0-nrgbd	NLL
Neural-RGBD with scale-invariant loss	ver0-nrgbd_silog	silog
STAD-frame (Ours)	ver1-per_frame_silog	silog
STAD (Ours)	ver4-aggr_silog	silog

Test on KITTI with given camera pose

In this case we assume the camera poses are given with the dataset.

In the code folder, run

sh local_test.sh

You need to change dataset_path and model_path correctly.

Contact

If you have any questions, please contact the author Hyunmin Lee<[e-mail]>.

Acknowledgement

Portions of the source code (e.g., training pipeline, argument parser, and logger) are from NVIDIA, Neural-RGBD