Need help with footprints?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

205 Stars 19 Forks Other 17 Commits 1 Opened issues


[CVPR 2020] Estimation of the visible and hidden traversable space from a single color image

Services available


Need anything else?

Contributors list

Footprints and Free Space from a Single Color Image

Jamie Watson, Michael Firman, Aron Monszpart and Gabriel J. Brostow – CVPR 2020 (Oral presentation)

[Link to Paper]

We introduce Footprints, a method for estimating the visible and hidden traversable space from a single RGB image

5 minute CVPR presentation video link

Understanding the shape of a scene from a single color image is a formidable computer vision task. Most methods aim to predict the geometry of surfaces that are visible to the camera, which is of limited use when planning paths for robots or augmented reality agents. Models which predict beyond the line of sight often parameterize the scene with voxels or meshes, which can be expensive to use in machine learning frameworks.

Our method predicts the hidden ground geometry and extent from a single image:

Web version of figure 1

Our predictions enable virtual characters to more realistically explore their environment.

Baseline exploration Our exploration
Baseline: The virtual character can only explore the ground visible to the camera Ours: The penguin can explore both the visible and hidden ground

⚙️ Setup

Our code and models were developed with PyTorch 1.3.1. The

list our dependencies.

We recommend installing and activating a new conda environment from these files with:

conda env create -f environment.yml -n footprints
conda activate footprints

🖼️ Prediction

We provide three pretrained models:

  • kitti
    , a model trained on the KITTI driving dataset with a resolution of 192x640,
  • matterport
    , a model trained on the indoor Matterport dataset with a resolution of 512x640, and
  • handheld
    , a model trained on our own handheld stereo footage with a resolution of 256x448.

We provide code to make predictions for a single image, or a whole folder of images, using any of these pretrained models. Models will be automatically downloaded when required, and input images will be automatically resized to the correct input resolution for each model.

Single image prediction:

python -m footprints.predict_simple --image test_data/cyclist.jpg --model kitti

Multi image prediction:

python -m footprints.predict_simple --image test_data --model handheld

By default,

predictions and
visualisations will be saved to the
folder; this can be changed with the

🚋 Training

To train a model you will need to download raw KITTI and Matterport data. Edit the

field in
to point to the downloaded raw data paths.

For details on downloading KITTI, see Monodepth2.

You will also need per-image training data generated from the video sequences: - visible ground segmentations - hidden ground depths - depth maps - etc.

Our versions of these can be found HERE. Download these and edit the

field of
to point to them.

After this your

should look like:
# Contents of paths.yaml

matterport: dataset: training_data:


Now you have everything you need to train!

Train a KITTI model using:

CUDA_VISIBLE_DEVICES=X python -m footprints.main \
    --training_dataset kitti \
    --log_path  \

and a Matterport model using:

CUDA_VISIBLE_DEVICES=X python -m footprints.main \
    --training_dataset matterport \
    --height 512  --width 640 \
    --log_path  \
    --batch_size 8 \

Training data generation

If you want to generate your own training data instead of using ours (e.g. you want to try a better ground segmentation algorithm, or more accurate camera poses) then you can!

There are several key elements of our training data - each can be swapped out for your own.

Visible depths

For KITTI we used PSMNet to generate disparity maps for stereo pairs. These are inside

, and are used to generate training labels. These are converted to depth maps using the known focal length and baseline. Matterport provides these.

Camera poses

For KITTI we used ORBSLAMv2 to generate camera poses, which are stored as

inside the
folder. These are used to reproject between cameras. Matterport provides these.

Ground segmentations

For both Matterport and KITTI we trained a segmentation network to classify ground pixels in an image. We provide training code for this inside

. These are stored inside the
folder as
and are unthresholded (i.e. raw sigmoid output).

Optical flow

For KITTI, we identify moving objects by comparing

induced flow
optical flow
. Our provided optical flow estimates come from LiteFlowNet, and are inside the

Hidden ground depths

To compute hidden depths (i.e. the depth to each visible and occluded ground pixel) we use camera poses, depth maps and ground segmentations. These can be generated using (expects a GPU to be available):

    footprints.preprocessing.ground_truth_generation.ground_truth_generator \
    --type hidden_depths  --data_type kitti --textfile splits/kitti/train.txt
Make sure to run this on both
. Warning - this will take a while, so to speed things up you can do this in parallel by running multiple processes and adding the flags
--start_idx X
--end_idx Y
to split the textfile into smaller chunks.

Note that if you have already downloaded our training data, running this command will overwrite it unless you set

. To actually train using this, you can manually set the path inside
, or rename your new data to the required folder name, e.g.

Moving object masks

To compute moving objects masks we use optical flow, depth, ground segmentations and camera poses. These can be generated by amending the above command with

--type moving_objects
. This is only valid for KITTI.

Depth masks

Depth masks are estimates of the untraversable pixels in the image, and are computed using depth maps and ground segmentations. To generate these change the above command to use

--type depth_masks

⏳ Evaluation

To generate predictions for evaluation using a trained model, run:

CUDA_VISIBLE_DEVICES=X python -m footprints.main \
    --mode inference \
    --load_path  \
    --inference_data_type  \
    --height <192 for kitti, 512 for matterport> \
    --width 640
By default this will save to
, but can be specified with

To evaluate a folder of predictions, run:

python -m footprints.evaluation.evaluate_model \
    --datatype kitti \
    --metric iou \

The following options are provided: -

can be either
. -
can be
) or

If necessary, the ground truth files will be automatically downloaded and placed in the


You can also download the KITTI annotations directly from here. For each image, there are 3

  • XXXXX_ground.png
    contains the mask of the boundary of visible and hidden ground, ignoring all objects
  • XXXXX_objects.png
    contains the mask of the ground space taken up by objects (the footprints)
  • XXXXX_combined.png
    contains the full evaluation mask - the visible and hidden ground, taking into account object footprints

Method and further results

We learn from stereo video sequences, using camera poses, per-frame depth and semantic segmentation to form training data, which is used to supervise an image-to-image network.

Video version of figure 3

Results on mobile phone footage:

Rig results Rig results

More results on the KITTI dataset:

KITTI results

✏️ 📄 Citation

If you find our work useful or interesting, please consider citing our paper:

 title   = {Footprints and Free Space from a Single Color Image},
 author  = {Jamie Watson and
            Michael Firman and
            Aron Monszpart and
            Gabriel J. Brostow},
 booktitle = {Computer Vision and Pattern Recognition ({CVPR})},
 year = {2020}

👩‍⚖️ License

Copyright © Niantic, Inc. 2020. Patent Pending. All rights reserved. Please see the license file for terms.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.