LoST? Appearance-Invariant Place Recognition for Opposite Viewpoints using Visual Semantics

This is the source code for the paper titled - "LoST? Appearance-Invariant Place Recognition for Opposite Viewpoints using Visual Semantics", [arXiv][RSS 2018 Proceedings]

An example output image showing Keypoint Correspondences:

An example output image showing Keypoint Correspondences

Flowchart of the proposed approach:

Flowchart of the proposed approach

If you find this work useful, please cite it as:
Sourav Garg, Niko Sunderhauf, and Michael Milford. LoST? Appearance-Invariant Place Recognition for Opposite Viewpoints using Visual Semantics. Proceedings of Robotics: Science and Systems XIV, 2018.

title={LoST? Appearance-Invariant Place Recognition for Opposite Viewpoints using Visual Semantics},
author={Garg, Sourav and Suenderhauf, Niko and Milford, Michael},
journal={Proceedings of Robotics: Science and Systems XIV},

RefineNet's citation as mentioned on their Github page.

Setup and Run


  • Ubuntu (Tested on 14.04)
  • RefineNet
    • Required primarily for visual semantic information. Convolutional feature maps based dense descriptors are also extracted from the same.
    • A modified fork of RefineNet's code is used in this work to simultaneously store convolutional dense descriptors.
    • Requires Matlab (Tested on 2017a)
  • Python (Tested on 2.7)
    • numpy (Tested on 1.11.1, 1.14.2)
    • scipy (Tested on 0.13.3, 0.17.1)
    • skimage (Minimum Required 0.13.1)
    • sklearn (Tested on 0.14.1, 0.19.1)
    • h5py (Tested on 2.7.1)
  • Docker (optional, recommended, tested on 17.12.0-ce)


  1. In your workspace, clone the repositories:
    git clone
    cd lostX
    git clone
    NOTE: If you download this repository as a zip, the refineNet's fork will not get downloaded automatically, being a git submodule.
  2. Download the Resnet-101 model pre-trained on Cityscapes dataset from here or here. More details on RefineNet's Github page.
    • Place the downloaded model's .mat file in the refinenet/model_trained/ directory.
  3. If you are using docker, download the docker image:
    docker pull souravgarg/vpr-lost-kc:v1


  1. Generate and store semantic labels and dense convolutional descriptors from RefineNet's conv5 layer In the MATLAB workspace, from the refinenet/main/ directory, run:


    The above will use the sample dataset from refinenet/datasets/ directory. You can set path to your data in demo_predict_mscale_cityscapes.m through variable datasetName and img_data_dir.
    You might have to run vl_compilenn before running the demo, please refer to the instructions for running refinenet in their official

  2. [For Docker users]
    If you have an environment with python and other dependencies installed, skip this step, otherwise run a docker container:

    docker run -it -v PATH_TO_YOUR_HOME_DIRECTORY/:/workspace/ souravgarg/vpr-lost-kc:v1 /bin/bash

    From within the docker container, navigate to lostX/lost_kc/ repository.
    -v option mounts the PATH_TO_YOUR_HOME_DIRECTORY to /workspace directory within the docker container.

  3. Reformat and pre-process RefineNet's output from lostX/lost_kc/ directory:


    $PATH_TO_REFINENET_OUTPUT is set to be the parent directory of predict_result_full, for example, ../refinenet/cache_data/test_examples_cityscapes/1-s_result_20180427152622_predict_custom_data/predict_result_1/

  4. Compute LoST descriptor:

  5. Repeat step 1, 3, and 4 to generate output for the other dataset by setting the variable datasetName to 2-s.

  6. Perform place matching using LoST descriptors based difference matrix and Keypoint Correspondences:

    python -n 10 -f 0 -p1 $PATH_TO_REFINENET_OUTPUT_1  -p2 $PATH_TO_REFINENET_OUTPUT_2

Note: Run python FILENAME -h for any of the python source files in Step 3, 4, and 6 for description of arguments passed to those files.


The code is released under MIT License.

Related Projects

Delta Descriptors (2020)

CoarseHash (2020)

seq2single (2019)

CRICOS No. 00213J