MobileStereoNet

This repository contains the code for "MobileStereoNet: Towards Lightweight Deep Networks for Stereo Matching", presented at WACV 2022 [Paper] [Supp] [arXiv] [Video Presentation].

Input image

2D-MobileStereoNet prediction

3D-MobileStereoNet prediction

Evaluation Results

MobileStereoNets are trained and tested using SceneFlow (SF), KITTI and DrivingStereo (DS) datasets.
In the following tables, the first columns show the training sets. For instance, in the case of "SF + KITTI2015", the model is firstly pretrained on the SceneFlow dataset, and then finetuned on KITTI images.
The results are reported in End-point Error (EPE); the lower, the better.
Note that some experiments evaluate the zero-shot cross-dataset generalizability, e.g. when the model is trained on "SF + DS" and evaluated on "KITTI2015 val" or "KITTI2012 train".
The related trained models are provided in the tables as hyperlinks.

• 2D-MobileStereoNet

	SF test	DS test	KITTI2015 val	KITTI2012 train
SF	1.14	6.59	2.42	2.45
DS	-	0.67	1.02	0.96
SF + DS	-	0.73	1.04	1.04
SF + KITTI2015	-	1.41	0.79	1.18
DS + KITTI2015	-	0.79	0.65	0.91
SF + DS + KITTI2015	-	0.83	0.68	0.90

• 3D-MobileStereoNet

	SF test	DS test	KITTI2015 val	KITTI2012 train
SF	0.80	4.50	10.30	9.38
DS	-	0.60	1.16	1.14
SF + DS	-	0.57	1.12	1.10
SF + KITTI2015	-	1.53	0.65	0.90
DS + KITTI2015	-	0.65	0.60	0.85
SF + DS + KITTI2015	-	0.62	0.59	0.83

Results on KITTI 2015 validation

Predictions of difference networks

Results on KITTI 2015 Leaderboard

Leaderboard
2D-MobileStereoNet on the leaderboard
3D-MobileStereoNet on the leaderboard

Computational Complexity

Requirements for computing the complexity by two methods:

pip install --upgrade git+https://github.com/sovrasov/flops-counter.pytorch.git
pip install --upgrade git+https://github.com/Lyken17/pytorch-OpCounter.git
pip install onnx

Run the following command to see the complexity in terms of number of operations and parameters.

python cost.py

You can also compute the complexity of each part of the network separately. For this, the input size of each module has been written in cost.py.

Installation

Requirements

The code is tested on:

• Ubuntu 18.04
• Python 3.6
• PyTorch 1.4.0
• Torchvision 0.5.0
• CUDA 10.0

Setting up the environment

conda env create --file mobilestereonet.yml
conda activate mobilestereonet

SceneFlow Dataset Preparation

Download the finalpass images and the disparity data for SceneFlow FlyingThings3D, Driving and Monkaa. For both, image and disparity data, move the directories in the TRAIN and TEST directories of the Driving and Monkaa Dataset (15mm_focallength/35mm_focallength for Driving, funnyworld_x2 etc. for Monkaa) into the FlyingThings3D TRAIN and TEST directories, respectively.

It should look like this:

frames_finalpass
│
└───TEST
│   │
│   └───A
│   └───B
│   └───C
│   
│
└───TRAIN
│   │
│   └───15mm_focallength
│   └───35mm_focallength
│   └───A
│   └───a_rain_of_stones_x2
│   └─── ..

Training

Set a variable for the dataset directory, e.g. DATAPATH="/Datasets/SceneFlow/". Then, run train.py as below:

Pretraining on SceneFlow

python train.py --dataset sceneflow --datapath $DATAPATH --trainlist ./filenames/sceneflow_train.txt --testlist ./filenames/sceneflow_test.txt --epochs 20 --lrepochs "10,12,14,16:2" --batch_size 8 --test_batch_size 8 --model MSNet2D

Finetuning on KITTI

python train.py --dataset kitti --datapath $DATAPATH --trainlist ./filenames/kitti15_train.txt --testlist ./filenames/kitti15_val.txt --epochs 400 --lrepochs "200:10" --batch_size 8 --test_batch_size 8 --loadckpt ./checkpoints/pretrained.ckpt --model MSNet2D

The arguments in both cases can be set differently depending on the model, dataset and hardware resources.

Prediction

The following script creates disparity maps for a specified model:

python prediction.py --datapath $DATAPATH --testlist ./filenames/kitti15_test.txt --loadckpt ./checkpoints/finetuned.ckpt --dataset kitti --colored True --model MSNet2D

Credits

The implementation of this code is based on PSMNet and GwcNet. Also, we would like to thank the authors of THOP: PyTorch-OpCounter, Flops counter and KITTI python utils.

License

This project is released under the Apache 2.0 license.

Citation

If you use this code, please cite this paper:

@inproceedings{shamsafar2022mobilestereonet,
  title={MobileStereoNet: Towards Lightweight Deep Networks for Stereo Matching},
  author={Shamsafar, Faranak and Woerz, Samuel and Rahim, Rafia and Zell, Andreas},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={2417--2426},
  year={2022}
}

Contact

The repository is maintained by Faranak Shamsafar.
[email protected]