Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



35 Commits

Repository files navigation

ISM: Invariant-based Stereo Matching

ISM is not a particular stereo matching algorithm. Rather, it is a generic algorithmic framework that accelerates any existing DNN-based stereo matching algorithm using what we call "stereo invariants". This reposity contains the software code that you can use to reproduce the accuracy results in our MICRO 2019 paper ASV: Accelerated Stereo Vision System.

We provide some implementation of ISM algorithm for several popular stereo DNNs.

  • PSMNet: "Pyramid Stereo Matching Network", CVPR 2017 [pdf]
  • FlowNet/DispNet: "FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks", CVPR 2018 [pdf]
  • GC-Net: "End-to-End Learning of Geometry and Context for Deep Stereo Regression", ICCV 2017 [pdf]

A generic framework for accelerating stereo vision

In the, it gives a skeleton code for using ISM with any stereo DNN model. It contains the main bone to implement our ISM algorithm to other stereo DNNs. There are several steps in TODO list in order to make this algorithm works.

First, you need to download a stereo vision dataset. For this, we recommend KITTI dataset and SceneFlow dataset and pass the path to the skeleton code.

Second, download any stereo vision DNN network to test this code. In the, you need to implement the functions load_dnn_model and dnn_inference to be able to load the stereo DNN model and use the stereo DNN model to geneerate disparity results from key frames.

Some details about this skeleton code

Inside of this skeleton script, we used OpticalFlowFarneback in OpenCV to compensate motions across adjacent frames. Other dense optical flow algrithm can also be used to substitute this function. The optical flow algorithm is used in one stage of ISM algorithm to get the disparity map from next subsequent frames.

Implementations of ISM with popular stereo vision DNNs

In the directory stereo_script, it contains several scripts that we applied our methods on four representative stereo DNNs. Here are some detailed instructions to run these scripts.


PSMNet is from a 2018 CVPR paper by Jia-Ren Chang and Yong-Sheng Chen.

To use our script, please first check out this link and follow the instructions to set up PSMNet model appropriately.

  $ git clone
  $ cd PSMNet

Next, you can download one of the stereo vision datasets. For the demostration purpose, we choose "SceneFlow". Download both "RGB images" and "Disparity" into the same directory PSMNet and untar them.

  $ tar -xzvf monkaa_frames_cleanpass.tar
  $ tar -xzvf monkaa_disparity.tar.bz2

After untaring both "RGB image" and "Disparity" tar, you will see two folders, frames_cleanpass and disparity, in the directory.

Next, copy two scripts from script/psmnet from our repository to their root directory, PSMNet. is a modified script that adds our ISM algorithm into PSMNet orignal python script. The is a bash script to test the ISM algorithm on different datasets in frames_cleanpass folder and place each test result into a folder named result.

A simple example of using

  $ python \
       --maxdisp 192 --model stackhourglass \
       --loadmodel pretrained_model_KITTI2015.tar \
       --saveimg False --datapath ${PATH} \
       --datasize 100

You need to specify the name of dataset as ${PATH} in the frames_cleanpass to run this python script. datasize flag specifies the number of images that you want to process in this dataset.

To test a set of dataset, you can run

  $ ./

You can also modified to specify the dataset that you want to test on. After finished running the test script, you can check all test results from result directory.

One of the example results is attached below:

Number of model parameters: 5224768
(1, 3, 544, 960) (1, 3, 544, 960)
old_disp:  (540, 960)
time = 39.79
Total index: 2, EW: 1
(1, 3, 544, 960) (1, 3, 544, 960)
pred_disp:  (540, 960)
time = 39.12
old_comparison:  24.032894 2.670288e-05 3.2993202 ...
esti_comparison:  31.811045 5.2452087e-06 3.6652005 ...
pred_comparison:  23.935379 2.2888184e-05 3.3056526 ...


For every pair of frames, the result shows the execution time to generate its disparity map, the index of image pair, and accuracy comparing with ground truth. There are two key comparison results:

  • esti_comparison : the accuracy using our ISM algorithm.
  • pred_comparison : the accuracy using DNN.

The first three fields in each comparison stand for:

  1. the maximum difference between ground truth and result;
  2. the minimum difference between ground truth and result;
  3. the average difference between ground truth and result;


To run FlowNet and DispNet along with our ISM algorithm, please first clone and setup Flownet2 from the paper by E. Ilg et al.

  $ git clone
  $ cd flownet2

Assuming now you are in the FlowNet root directory, after you get FlowNet running successfully, copy the two scripts from script/flownet to their ./scripts directory.

And then, download the "SceneFlow" dataset to the root directory, flownet2, and untar the dataset.

  $ tar -xzvf monkaa_frames_cleanpass.tar
  $ tar -xzvf monkaa_disparity.tar.bz2

To run the ISM algorithm with FlowNet-C, run the command below:

  $ ./ \
         ../models/FlowNet2-c/FlowNet2-c_weights.caffemodel \
         ../models/FlowNet2-c/FlowNet2-c_deploy.prototxt.template \
         --ew 4 --path ${PATH}

You can switch to DispNet by replace FlowNet2-c to FLowNet2-S.

You need to specify the path to the dataset, in this case, it is the root directory of flownet2.

There is also a batch-processing Bash script you can use to test on a set of datasets. Just simply run:

  $ ./

Then, you can check out results from result directory as previously shown. The format of result is the same as the one in PSMNet.


GC-Net is from a 2017 CVPR paper: End-to-End Learning of Geometry and Context for Deep Stereo Regression by A. Kendall et al. We used the TensorRT version from NVDIA. To check out the stereo DNN, please to go this link. Clone this repo:

  $ git clone 

And then, redirect to the redtail/stereoDNN directory.

  $ cd redtail/stereoDNN

And follow the instructions to build the inference code (release version). After build succeeds, you should see a nvstereo binary in bin directory.

Now, you can make a new directory in the root directory to test our ISM algorithm.

  $ mkdir ism

Then, copy both binary and its weights into this directory.

  $ cp bin/nvstereo ism/nvstereo
  $ cp models/ResNet-18/TensorRT/trl_weights.bin ism/trl_weights.bin

Next, you can follow the previous instruction to download the dataset. For the demostration purpose, we choose "SceneFlow". Download both "RGB images" and "Disparity" in the same directory PSMNet and untar them.

  $ tar -xzvf monkaa_frames_cleanpass.tar
  $ tar -xzvf monkaa_disparity.tar.bz2

After untaring both "RGB image" and "Disparity" tar, you will see two folders, frames_cleanpass and disparity, in the directory.

Last, copy two script from our directory, stereo_script/gcnet, to ism directory. Now, you can run our ism algorithm using:

$ ./ \
       --ew 4 --path ${PATH}

The ${PATH} is the path lead to one particular dataset.

You can also use our script test_dataset to run multiple dataset together.


This project is a artifact of our 2019 MICRO paper:

Y. Feng, P. Whatmough, and Y. Zhu, "ASV: Accelerated Stereo Vision System", In Proc. of MICRO, 2019.

Please kindly consider citing this paper in your publications if it helps your research.

  title={ASV: Accelerated Stereo Vision System},
  author={Feng, Yu and Whatmough, Paul and Zhu, Yuhao},
  booktitle={Proceedings of the 52th International Symposium on Microarchitecture},


If there is any question, please email me:


A invariant-based stereo matching algorithm







No releases published


No packages published