BTS

From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation
arXiv
Supplementary material

Video Demo 1

Video Demo 2

Note

This repository contains a Tensorflow implementation of BTS.
We tested this code under python 2.7 and 3.6, Tensorflow 1.14.0, CUDA 10.0 on Ubuntu 18.04.
If you use TensorFlow built from source, it is okay with v1.14.0.
If you use TensorFlow installed using pip, it is okay up to v1.13.2.
Currently, if we use TensorFlow v1.14.0 installed using pip, we get segmentation fault.

Preparation

$ cd ~
$ mkdir workspace
$ cd workspace
$ git clone https://github.com/cogaplex-bts/bts
$ cd bts/custom_layer
$ mkdir build && cd build
$ cmake -D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda ..
$ make -j

For TensorFlow v1.14.0

If you encounter an error "fatal error: third_party/gpus/cuda/include/cuda_fp16.h: No such file or directory", open "tensorflow/include/tensorflow/core/util/gpu_kernel_helper.h" and edit a line from

#include "third_party/gpus/cuda/include/cuda_fp16.h"

to

#include "cuda_fp16.h"

Also, you will need to edit lines in "tensorflow/include/tensorflow/core/util/gpu_device_functions.h" from

#include "third_party/gpus/cuda/include/cuComplex.h"
#include "third_party/gpus/cuda/include/cuda.h"

to

#include "cuComplex.h"
#include "cuda.h"

For TensorFlow up to v1.13.2

If you are testing with Tensorflow version lower than 1.14, please edit a line in "compute_depth.cu" from

#include "tensorflow/include/tensorflow/core/util/gpu_kernel_helper.h"

to

#include "tensorflow/include/tensorflow/core/util/cuda_kernel_helper.h"

Then issue the make commands again.

$ cmake ..
$ make -j

Testing with NYU Depth V2

$ cd ~/workspace/bts/utils
# Get official NYU Depth V2 split file
$ wget http://horatio.cs.nyu.edu/mit/silberman/nyu_depth_v2/nyu_depth_v2_labeled.mat
# Convert mat file to image files
$ python extract_official_train_test_set_from_mat.py nyu_depth_v2_labeled.mat splits.mat ../../dataset/nyu_depth_v2/official_splits/
$ cd ..
$ mkdir models
# Get BTS model trained with NYU Depth V2
$ python utils/download_from_gdrive.py 1ipme-fkV4pIx87sOs31R9CD_Qg-85__h models/bts_nyu.zip
$ cd models
$ unzip bts_nyu.zip

Once the preparation steps completed, you can test BTS using following commands.

$ cd ~/workspace/bts
$ python bts_test.py arguments_test_nyu.txt

This will save results to ./result_bts_nyu. With a single RTX 2080 Ti it takes about 34 seconds for processing 654 testing images.

Evaluation

Following command will evaluate the prediction results for NYU Depvh V2.

$ python eval_with_pngs.py --pred_path ./result_bts_nyu/raw/ --gt_path ../dataset/nyu_depth_v2/official_splits/test/ --dataset nyu --min_depth_eval 1e-3 --max_depth_eval 10 --eigen_crop

You should see outputs like this:

Raw png files reading done
Evaluating 654 files
GT files reading done
0 GT files missing
Computing errors
     d1,      d2,      d3,  AbsRel,   SqRel,    RMSE, RMSElog,   SILog,   log10
  0.882,   0.980,   0.996,   0.111,   0.060,   0.356,   0.141,  11.268,   0.047
Done.

Preparation for Training

NYU Depvh V2

First, you need to download DenseNet-161 model pretrained with ImageNet.

# Get DenseNet-161 model pretrained with ImageNet
$ cd ~/workspace/bts
$ python utils/download_from_gdrive.py 1rn7xBF5eSISFKL2bIa8o3d8dNnsrlWfJ models/densenet161_imagenet.zip
$ cd models && unzip densenet161_imagenet.zip

Then, download the dataset we used in this work.

$ cd ~/workspace/bts
$ python utils/download_from_gdrive.py 1AysroWpfISmm-yRFGBgFTrLy6FjQwvwP ../dataset/nyu_depth_v2/sync.zip
$ unzip sync.zip

Or, you can prepare the dataset by yourself using original files from official site NYU Depth V2. There are two options for downloading original files: Single file downloading and Segmented-files downloading.

Single file downloading:

$ cd ~/workspace/dataset/nyu_depth_v2
$ mkdir raw && cd raw
$ wget http://horatio.cs.nyu.edu/mit/silberman/nyu_depth_v2/nyu_depth_v2_raw.zip
$ unzip nyu_depth_v2_raw.zip

Segmented-files downloading:

$ cd ~/workspace/dataset/nyu_depth_v2
$ mkdir raw && cd raw
$ aria2c -x 16 -i ../../../bts/utils/nyudepthv2_archives_to_download.txt
$ cd ~/workspace/bts
$ python utils/download_from_gdrive.py 1xBwO6qU8UCS69POJJ0-9luaG_1pS1khW ../dataset/nyu_depth_v2/raw/bathroom_0039.zip
$ python utils/download_from_gdrive.py 1IFoci9kns6vOV833S7osV6c5HmGxZsBp ../dataset/nyu_depth_v2/raw/bedroom_0076a.zip
$ python utils/download_from_gdrive.py 1ysSeyiOiOI1EKr1yhmKy4jcYiXdgLP4f ../dataset/nyu_depth_v2/raw/living_room_0018.zip
$ python utils/download_from_gdrive.py 1QkHkK46VuKBPszB-mb6ysFp7VO92UgfB ../dataset/nyu_depth_v2/raw/living_room_0019.zip
$ python utils/download_from_gdrive.py 1g1Xc3urlI_nIcgWk8I-UaFXJHiKGzK6w ../dataset/nyu_depth_v2/raw/living_room_0020.zip
$ parallel unzip ::: *.zip

Get official toolbox for rgb and depth synchronization.

$ cd ~/workspace/bts/utils
$ wget http://cs.nyu.edu/~silberman/code/toolbox_nyu_depth_v2.zip
$ unzip toolbox_nyu_depth_v2.zip
$ cd toolbox_nyu_depth_v2
$ mv ../sync_project_frames_multi_threads.m .
$ mv ../train_scenes.txt .

Run script "sync_project_frames_multi_threads.m" using MATLAB to get synchronized RGB and depth images. This will save rgb-depth pairs in "~/workspace/dataset/nyu_depth_v2/sync". Once the dataset is ready, you can train the network using following command.

$ cd ~/workspace/bts
$ python bts_main.py arguments_train_nyu.txt

You can check the training using tensorboard:

$ tensorboard --logdir ./models/bts_nyu_test/

Open localhost:6006 with your favorite browser to see the progress of training.

KITTI

You can also train BTS with KITTI dataset by following procedures. First, download the ground truth depthmaps from KITTI. Then, download and unzip the raw dataset using following commands.

$ cd ~/workspace/dataset
$ mkdir kitti_dataset && cd kitti_dataset
$ mv ~/Downloads/data_depth_annotated.zip .
$ unzip data_depth_annotated.zip
$ aria2c -x 16 -i ../../bts/utils/kitti_archives_to_download.txt
$ parallel unzip ::: *.zip

Finally, we can train our network with

$ cd ~/workspace/bts
$ python bts_main.py arguments_train_eigen.txt

Testing and Evaluation with KITTI

Once you have KITTI dataset and official ground truth depthmaps, you can test and evaluate our model with following commands.

# Get KITTI model trained with KITTI Eigen split
$ cd ~/workspace/bts
$ python utils/download_from_gdrive.py 1w4WbSQxui8GTDEsjX5xb4m7_-5yCznhQ models/bts_eigen.zip
$ cd models && unzip bts_eigen.zip

Test and save results.

$ cd ~/workspace/bts
$ python bts_test.py arguments_test_eigen.txt

This will save results to ./result_bts_eigen. Finally, we can evaluate the prediction results with

$ python eval_with_pngs.py --pred_path ./result_bts_eigen/raw/ --gt_path ../dataset/kitti_dataset/data_depth_annotated/ --dataset kitti --min_depth_eval 1e-3 --max_depth_eval 80 --do_kb_crop --garg_crop

You should see outputs like this:

GT files reading done
45 GT files missing
Computing errors
     d1,      d2,      d3,  AbsRel,   SqRel,    RMSE, RMSElog,   SILog,   log10
  0.951,   0.993,   0.998,   0.064,   0.256,   2.796,   0.100,   9.175,   0.028
Done.

License

Copyright (C) 2019 Jin Han Lee, Myung-Kyu Han, Dong Wook Ko and Il Hong Suh
This Software is licensed under GPL-3.0-or-later.