SMALify
This repository contains an implementation for performing 3D animal (quadruped) reconstruction from a monocular image or video. The system adapts the pose (limb positions) and shape (animal type/height/weight) parameters for the SMAL deformable quadruped model, as well as camera parameters until the projected SMAL model aligns with 2D keypoints and silhouette segmentations extracted from the input frame(s).
The code can be thought of as a modernization of the fitting code used in Creatures Great and SMAL paper; Chainer/ChumPy has been replaced with PyTorch, OpenDR replaced with PyTorch3D etc. However, I have also included some recent innovations from the Who Left the Dogs Out? such as the inclusion of limb scaling parameters, and an improved shape prior.
The aim of this repository is to provided demonstrative fitting code to benefit computer vision researchers but also those working in animal/veterinary science. In either case, I'd be delighted to hear from you!
Installation
-
Clone the repository with submodules and enter directory
git clone --recurse-submodules https://github.com/benjiebob/SMALify cd SMALify
Note: If you don't clone with submodules you won't get the sample data from BADJA/StanfordExtra/SMALST.
-
Install dependencies, particularly PyTorch (cuda support recommended), Pytorch3D. Check requirements.txt for full details.
-
Download images corresponding to demo BADJA sequence
cd data/BADJA wget http://mi.eng.cam.ac.uk/~bjb56/datasets/badja_extra_videos.zip unzip badja_extra_videos.zip
-
Inspect the directory paths in config.py and make sure they match your system.
QuickStart: Running the Fitter
- Run on a sample video from the BADJA dataset.
-
Run the python script
python smal_fitter/optimize_to_joints.py
-
Inspect the SMALify/checkpoint directory to inspect the progress of the fitter.
-
The output files stX_epY means, stage X and iteration Y of the fitter.
-
The final output is named st10_ep0 by (a slightly lazy) convention.
-
The other files have the following meaning
Extension Explanation .png Image Visualization .ply Mesh file, can be viewed in e.g. MeshLab .pkl Pickle file, contains the latest model/camera parameters
-
-
Create a video with the final fits
- The generate_video.py function loads the exported .pkl files generated during the fitting process and exports the data. This is generally usful if your .pkl files are created using alternative methods, e.g. Who Left the Dogs Out? (coming soon!) or your own research.
- Set CHECKPOINT_NAME in config.py to be the name of the output directory in SMALify/checkpoints.
- By default the code will load the final optimized meshes, indicated by EPOCH_NAME = "st10_ep0". If you want to generate a video from intermediate results, set this to some different stage/iteration.
- Run the video generation script, which exports the video to SMALify/exported
python smal_fitter/generate_video.py
- Create a video using e.g. FFMPEG:
cd exported/CHECKPOINT_NAME/EPOCH_NAME ffmpeg -framerate 2 -pattern_type glob -i '*.png' -pix_fmt yuv420p results.mp4
-
- Fit to an image from StanfordExtra dataset.
- Edit the config.py file to make load a StanfordExtra image instead of a BADJA video sequence:
# SEQUENCE_OR_IMAGE_NAME = "badja:rs_dog" SEQUENCE_OR_IMAGE_NAME = "stanfordextra:n02099601-golden_retriever/n02099601_176.jpg"
- Run the python script:
python smal_fitter/optimize_to_joints.py
- Edit the config.py file to make load a StanfordExtra image instead of a BADJA video sequence:
Running on alternative data
Alternative BADJA/StanfordExtra sequences:
-
Follow the instructions for BADJA or StanfordExtra.
-
Open the config.py file and make the following changes
Config Setting Explanation Example SEQUENCE_OR_IMAGE_NAME Used to refer to your sequence/image badja:rs_dog SHAPE_FAMILY Choose from 0: Cat, 1: Canine (e.g. Dog), 2: Equine (e.g. Horse), 3: Bovine (e.g. Cow), 4: Hippo 1 IMAGE_RANGE Number of frames to process from the sequence. Ignored for StanfordExtra. [1,2,3] or range(0, 10) WINDOW_SIZE For video sequences, the number of frames to fit into a batch. Alter depending on GPU capacity. 10
Running on your own data (TODO... Ben working on this now!)
The first job is to generate keypoint/silhouette data for your input image(s).
- Manual Annotation
- I recommend using LabelMe, which is fantastic software that makes annotating keypoints / silhouettes efficient.
- Install the software, and then load the joint annotation execute
labelme --labels labels.txt --nosortlabels
- Next, generate the silhouette annotations
# TODO
- TODO: Write script to load labelme files
- Automatic Dog Joint Predictor
- TODO: Release model for dog keypoint prediction, trained on StanfordExtra.
- Automatic Quadruped Joint Predictor
- TODO: Release training/testing scripts for CreaturesGreatAndSMAL joint prediction, trained on synthetic SMAL renders, OJA methods etc.
Building your own quadruped deformable model
If you want to represent an animal quadruped category which isn't covered by the SMAL model (e.g. perhaps you want to reconstruct rodents/squirrels), you can use the fitter_3d tool. The basic idea is to fit the existing SMAL model to a collection of 3D artist meshes (you can download online) and thereby learn a new shape space. More information is given in the README.
Improving performance and general tips and tricks
- For some sequences, it may be necessary to fiddle with the weights applied to each part of the loss function. These are defined in config.py in an OPT_WEIGHTS settings. The values weight the following loss components:
Loss Component | Explanation | Tips for Loss Weight |
---|---|---|
2D Keypoint Reprojection | Project the SMAL model with latest parameters and compare projected joints to input 2D keypoints | If your model limbs don't match well with the input keypoints after fitting, it may be worth increasing this. |
3D Shape Prior | Used to constrain the 3D shapes to be 'animal-like'. Note that (unlike equivalent human approaches that use mocap etc.) only artist data is used for this. | If your reconstructed animals don't look like animals, try increasing this. |
3D Pose Prior | Used to contains the 3D poses to be anatomically plausible. | If your reconstructed animals have limb configurations which are very unreasonable, e.g. legs in strange places, try increasing this. |
2D Silhouette | Project the SMAL model with latest parameters and compare rendered silhouette to input 2D silhouette. | If the shape of your reconstructed animal doesn't match well (e.g. maybe it's too thin?), try increasing this. |
Temporal | Constrain the change in SMAL parameters between frames. (Only for videos) | If your limbs move unnaturally between video frames, try adapting this. |
Note that to avoid poor local minima, the optimization proceeds over multiple stages and the weights vary at each stage. For example, only once an approximate camera location has been found should there be 2D joint loss, and only once an approximate set of limb positions have been found should there be a 2D silhouette loss.
- To improve efficiency it is very likely that the number of iterations (again a setting in OPT_WEIGHTS) can be reduced for many sequences. I've err'd on the side of caution for this release by running many more iterations than probably needed.
Acknowledgements
This repository owes a great deal to the following works and authors:
- SMAL; Zuffi et al. designed the SMAL deformable quadruped template model and have been wonderful for providing advice throughout my animal reconstruction PhD journey.
- SMPLify; Bogo et al. provided the basis for our original ChumPY implementation and inspired the name of this repo.
- SMALST; Zuffi et al. provided a PyTorch implementations of the SMAL skinning functions which have been used here.
If you find this fitting code and/or BADJA dataset useful for your research, please consider citing the following paper:
@inproceedings{biggs2018creatures,
title={{C}reatures great and {SMAL}: {R}ecovering the shape and motion of animals from video},
author={Biggs, Benjamin and Roddick, Thomas and Fitzgibbon, Andrew and Cipolla, Roberto},
booktitle={ACCV},
year={2018}
}
if you make use of the limb scaling parameters, or Unity shape prior (on by default for the dog shape family) or the StanfordExtra dataset please cite Who Left the Dogs Out? 3D Animal Reconstruction with Expectation Maximization in the Loop:
@inproceedings{biggs2020wldo,
title={{W}ho left the dogs out?: {3D} animal reconstruction with expectation maximization in the loop},
author={Biggs, Benjamin and Boyne, Oliver and Charles, James and Fitzgibbon, Andrew and Cipolla, Roberto},
booktitle={ECCV},
year={2020}
}
Contribute
Please create a pull request or submit an issue if you would like to contribute.
Licensing
(c) Benjamin Biggs, Oliver Boyne, Andrew Fitzgibbon and Roberto Cipolla. Department of Engineering, University of Cambridge 2020
By downloading this code, you agree to the Creative Commons Attribution 3.0 International license. This license allows users to use, share and adapt the code, so long as credit is given to the authors (e.g. by citation).
THIS SOFTWARE AND ANNOTATIONS ARE PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.