Learning Augmentation Policy Schedules for Unsuperivsed Depth Estimation

Setup

Assuming a fresh Anaconda distribution, you can install the dependencies with:

pip install ray
pip install numpy==1.16.1
pip install comet_ml
pip install torch==1.5.0+cu92 torchvision==0.6.0+cu92 -f https://download.pytorch.org/whl/torch_stable.html
pip install requests

We ran our experiments with PyTorch-gpu 1.5, CUDA 9.1, Python 3.5, TensorFlow-gpu 1.11, Ray 0.8.4, Cudnn7 and Ubuntu 18.04.

You can pull the docker image with the above setup from : amanraj42/tensorflow-gpu-ray:v1.11 . We also provide a docker file in this repository.

Checkpoints

Our checkpoints are available at : https://drive.google.com/drive/u/3/folders/19QAZam6TdoRyTdCjbdoripW4xldp4ZV2 Method names are in Table 4.2 and Table 4.3 in the thesis.

Dataset preparation

Follow the Dataset Preparation instuctions from GeoNet.

Prediction for a single image

Download the checkpoints and use the one interested with:

python test_depth_simple.py --scale_normalize --image_path path/to/image.jpg --ckpt_path path/to/model.ckpt-*

Search : Learning Data Augmentation Policy Schedules

Have a look at bash script for policy search schedule using PBA located at scripts/search.sh in the repository. You can point to relevant data paths and execute it by sh scripts/search.sh We also provide policy schedules with which we trained our methods. They are named with method name as in Table 4.3 in the thesis.

Our policy schedules were found by PBA search on a reduced version of KITTI Eigen train set, that contained 5000 frame triplets. This is provided in data/test_files_eigen.txt

Training : Full scale training on KITTI

After running search, one can use generated policy schedules for training on full dataset of KITTI Eigen split. Refer to script scripts/train_mono2.sh for details on training arguments.

For validation we use KITTI eigen test set which is provided in data/test_files_eigen.txt

Result

Citation

If you use this work in your research, please cite:

Raj, A. (2020). Learning Augmentation Policy Schedules for Unsuperivsed Depth Estimation. UC San Diego. ProQuest ID: Raj_ucsd_0033M_19587. Merritt ID: ark:/13030/m5ns636d. Retrieved from https://escholarship.org/uc/item/1p85x50q

Acknowledgement

This code is based on the excellent repo by PBA folks at: https://github.com/arcelien/pba

amanrajdce/laps-depth