GANet

GA-Net: Guided Aggregation Net for End-to-end Stereo Matching

Brief Introduction

We are formulating traditional geometric and optimization of stereo into deep neural networks ...

Oral Presentation

Slides, Video, Poster

Building Requirements:

gcc: >=5.3
GPU mem: >=6.5G (for testing);  >=11G (for training, >=22G is prefered)
pytorch: >=1.0
cuda: >=9.2 (9.0 doesn’t support well for the new pytorch version and may have “pybind11 errors”.)
tested platform/settings:
  1) ubuntu 16.04 + cuda 10.0 + python 3.6, 3.7
  2) centos + cuda 9.2 + python 3.7
  

Install Pytorch:

You can easily install pytorch (>=1.0) by "pip install" to run the code. See this https://github.com/feihuzhang/GANet/issues/24

But, if you have trouble (lib conflicts) when compiling cuda libs, installing pytorch from source would help solve most of the errors (lib conflicts).

Please refer to https://github.com/pytorch/pytorch about how to reinstall pytorch from source.

How to Use?

Step 1: compile the libs by "sh compile.sh"

• Change the environmental variable ($PATH, $LD_LIBRARY_PATH etc.), if it's not set correctly in your system environment (e.g. .bashrc). Examples are included in "compile.sh".
• If you met the BN error, try to replace the sync-bn with another version:
  1. Install NVIDIA-Apex package https://github.com/NVIDIA/apex $ git clone https://github.com/NVIDIA/apex $ cd apex $ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
  2. Revise the "GANet_deep.py": add import apex change all BatchNorm2d and BatchNorm3d to apex.parallel.SyncBatchNorm

Step 2: download and prepare the dataset

download SceneFLow dataset: "FlyingThings3D", "Driving" and "Monkaa" (final pass and disparity files).

  -mv all training images (totallty 29 folders) into ${your dataset PATH}/frames_finalpass/TRAIN/
  -mv all corresponding disparity files (totallty 29 folders) into ${your dataset PATH}/disparity/TRAIN/
  -make sure the following 29 folders are included in the "${your dataset PATH}/disparity/TRAIN/" and "${your dataset PATH}/frames_finalpass/TRAIN/":
    
    15mm_focallength	35mm_focallength		A			 a_rain_of_stones_x2		B				C
    eating_camera2_x2	eating_naked_camera2_x2		eating_x2		 family_x2			flower_storm_augmented0_x2	flower_storm_augmented1_x2
    flower_storm_x2	funnyworld_augmented0_x2	funnyworld_augmented1_x2	funnyworld_camera2_augmented0_x2	funnyworld_camera2_augmented1_x2	funnyworld_camera2_x2
    funnyworld_x2	lonetree_augmented0_x2		lonetree_augmented1_x2		lonetree_difftex2_x2		  lonetree_difftex_x2		lonetree_winter_x2
    lonetree_x2		top_view_x2			treeflight_augmented0_x2	treeflight_augmented1_x2  	treeflight_x2	

download and extract kitti and kitti2015 datasets.
    

Step 3: revise parameter settings and run "train.sh" and "predict.sh" for training, finetuning and prediction/testing. Note that the “crop_width” and “crop_height” must be multiple of 48, "max_disp" must be multiple of 12 (default: 192).

Pretrained models:

• These pre-trained models use a batchsize of 8 on four P40 GPUs with a crop size of 240x624.
• Eight 1080ti/Titan GPUs should also be able to achieve the similar accuracy.
• Eight P40/V100/Titan RTX (22G) GPUs would be even better.

Results:

The results of the deep model are better than those reported in the paper.

Evaluations and Comparisons on SceneFlow Dataset (only 10 epoches)

Evaluations on KITTI 2012 benchmark

Evaluations on KITTI 2015 benchmark

Great Generalization Abilities:

GANet has great generalization abilities on other datasets/scenes.

Cityscape

Middlebury

Reference:

If you find the code useful, please cite our paper:

@inproceedings{Zhang2019GANet,
  title={GA-Net: Guided Aggregation Net for End-to-end Stereo Matching},
  author={Zhang, Feihu and Prisacariu, Victor and Yang, Ruigang and Torr, Philip HS},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={185--194},
  year={2019}
}