This repository contains the unofficial PyTorch implementation of From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation
Paper
Official Tensorflow Implementation
This repository is tested on Windows and Ubuntu on PyTorch 1.2, 1.3 and 13.1, installed from pip and built from source.
Kitti Validation results:
Model | Silog | rmse | rmse log | abs relative | sqrt relative |
---|---|---|---|---|---|
BTS - PyTorch | 9.83 | 3.03 | 0.10 | 0.06 | 0.29 |
BTS - Official | 9.08 | 2.82 | 0.09 | 0.05 | 0.26 |
As can be seen on table above, this implementation performs slightly worse than original implementation, which is very likely due some additional hyper-parameter tuning done by authors, due computational reasons I couldnt fine tune training parameters further.
pip install -r requirements.txt
Download pretrained model and put it under models directory
Please refer to prediction_example.ipynb
Kitti: Preperation process is same as the official tensorflow implementation. But use "kitti_archives_to_download.txt" provided in this reposity which contains more runs.
Change following lines at the start of the configs.py
model_path = "models/btspytorch"
dataset_path = "e://Code/Tez/bts_eren/kitti"
Run Test.py
Change following lines at the start of the configs.py
experiment_name = "Balatkan" # This determines folder names used for saving tensorboard logs and model files
dataset_path = "e://Code/Tez/bts_eren/kitti"
Takes around 100 hours to train on GTX 1080.