DynaDepth

This is the official PyTorch implementation for [Towards Scale-Aware, Robust, and Generalizable Unsupervised Monocular Depth Estimation by Integrating IMU Motion Dynamics], ECCV2022

If you find this work useful in your research, please consider citing our paper:

@inproceedings{zhang2022towards,
  title={Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating IMU motion dynamics},
  author={Zhang, Sen and Zhang, Jing and Tao, Dacheng},
  booktitle={European Conference on Computer Vision},
  pages={143--160},
  year={2022},
  organization={Springer}
}

Method Overview

Results on KITTI

Generalization on Make3D

Data Preparation

1. Download both the raw (unsync) and the sync kitti datasets from https://www.cvlibs.net/datasets/kitti/raw_data.php. For each sequence, you will have two folders XXX_extract/ and XXX_sync, e.g. 2011_10_03/2011_10_03_drive_0042_extract and 2011_10_03/2011_10_03_drive_0042_sync
2. The experiments are performed using the data from the sync kitti dataset (XXX_sync/). Since the imu (oxt/) in the sync dataset is sampled at the same frequency of the images, we need to perform a matching preprocessing step using the imu data in the raw dataset to get the corresponding imu data at the original frequency.

• You can achieve this by using python match_kitti_imu.py
• What you need to do: (1) Modify line 71-76 to get the sequence names of your own setting (2) Modify line 89-90 to your own path to the raw and the sycn datasets
• The matched results will be saved in matched_oxts\ under each sequence folder XXX_sync
• A 5ms drift is allowed for current matching process. You can modify line 153 if you are not happy about this setting
• Note that we directly match the imu data using the timestamps, while ignoring potential time asynchronization between the imu and the camera timing systems.

3. Since the unsync dataset is quite large to download, we also provide our preprocessed imu files in the following link: https://pan.baidu.com/s/1971KrQEHw5kVRy_Y4Lj5FA pwd:80pz

4. For the image preprocessing, we follow the practice in https://github.com/nianticlabs/monodepth2 to convert the image format from png to jpg for a smaller image size:

find kitti_data/ -name '*.png' | parallel 'convert -quality 92 -sampling-factor 2x2,1x1,1x1 {.}.png {.}.jpg && rm {}'

5. Since I only did preprocessing for once at the beginning of this project, please remind me by raising a new issue if I miss anything here

Training

This codebase is developed under PyTorch-1.4.0, CUDA-10.0, and Ubuntu-18.04.1.

You can train our full model with:

python train.py --data_path YOUR_PATH_TO_DATA --use_ekf --num_layers 50

To use ResNet-18 rather than ResNet-50 as the backbone, you can change --num_layer to 18

To disable the ekf fusion and use the IMU-related losses only, you can simply remove --use_ekf

To use loss weights other than the default setting, you can manipulate with the options, e.g.,

• --imu_warp_weight 0.5 --imu_consistency_weight 0.01
• --velo_weight 0.001 --gravity_weight 0.001

Evaluation

You can evaluate on the KITTI test set with:

python evaluate_depth.py --num_layer 50 --load_weights_folder YOUR_PATH_TO_MODEL_WEIGHTS --post_process

By default, we report the learnt scale without the median scaling trick. Use --eval_mono if you want to test the performance with median scaling

For evaluation without post processing, simply remove --post_process.

To evaluate the models with ResNet-18 backbone, change --num_layer to 18 accordingly.

To evaluate the models on Make3D, use evaluate_make3d.py with the same arguments as evaluate_depth.py. But you need to change the variable main_path in read_make3d() to your own path that contains test images of Make3D.

Our pretrained models

The full pretrained models corresponding to the results in our ECCV paper can be downloaded from the following links:

DynaDepth R18: https://pan.baidu.com/s/1ksP2m-6rQ_PkBTLmjAAuLQ pwd:xc5h

DynaDepth R50: https://pan.baidu.com/s/1X7OAOKFZ4fw3crOx6bn4ZA pwd:c3kj

Acknowledgment

This repo is built upon the excellent works of monodepth2, deep_ekf_vio, and liegroups. The borrowed codes are licensed under their original license respectively.