This repository is based on SMALify with the aim to turn any rigged 3D model into a SMAL compatible model. There are Blender files to convert your mesh and lots of code changes to deal with arbitrary armature configurations, rather than assuming a fixed quadruped model.
For now, I'll focus on insects, hence SMIL.
Important
Produced poses may be unrealistic, as we currently don't have a learned pose prior and are instead working with user-defined joint limits (currently set in smal_fitter/priors/joint_limits_prior.py but soon to be adjustable interactively in the accompanying blender addon)
-
Clone the repository with submodules and enter directory
git clone --recurse-submodules https://github.com/FabianPlum/SMILify
Note: If you don't clone with submodules you won't get the sample data from BADJA/StanfordExtra/SMALST.
-
install pytorch (and Co.)
conda create -n pytorch3d python=3.10 conda activate pytorch3d conda install pytorch=2.3.1 torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia conda install -c conda-forge -c fvcore iopath ninja imageio scikit-image pip install yacs pycocotools pip install --upgrade iopath
-
clone pytorch3d and install (WINDOWS)
git clone https://github.com/facebookresearch/pytorch3d.git cd pytorch3d pip install -e . cd ..
on LINUX just run
conda install pytorch3d -c pytorch3d
-
some more dependencies
pip install matplotlib scipy chumpy opencv-python nibabel trimesh
-
Test your installation
pytest tests/pipeline_tests.py -v -s
In case you get an error back with chumpy complaining about legacy imports and not finding numpy.bool, simply update the init.py file of chumpy:
replace
from numpy import bool, int, float, ...
with
from numpy import bool_ as bool
from numpy import int_ as int
from numpy import complex_ as complex
from numpy import unicode_ as unicode
from numpy import str_ as str
from numpy import float_ as float
from numpy import object_ as object
from numpy import inf as inf
from numpy import nan as nan
-
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 BADJA videos and unzip to
badja_extra_videos.zip
. -
Inspect the directory paths in config.py and make sure they match your system.
- Run on a synthetic sample image generated with replicAnt.
-
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:
-
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
The first job is to generate keypoint/silhouette data for your input image(s). 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
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.
- 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.
- Remove all currently used recursive clones. The repo should work on its own without the need of cloning submodules.
- If a submodule is needed, we should re-write it and add it to an appropriate subfolder. Otherwise, this repo is entirely un-maintainable.
- At the moment, the SMAL models require 2 to 3 separate types of data files as well as hard-coded priors for the joint limits. These should be handled more gracefully, like in the new SMIL implementation. All model info should be contained in a single, readable and editable file.
- Get rid of the numpy/chumpy dependency mess.
- Write a conversion script from the old SMAL format consisting of multiple files into our new single file structure containing all the data. I don't care if the files are large, as long as they are readable and first and foremost editable.
- The code is poorly documented. That needs to be fixed.
- The code is poorly tested. That needs to be fixed.
- Let's see how far we can get with this in our limited time BUT I would love to re-write this whole thing as a Blender addon. But that's for another day.
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}
}
Please create a pull request or submit an issue if you would like to contribute.
(c) Benjamin Biggs, Oliver Boyne, Andrew Fitzgibbon and Roberto Cipolla. Department of Engineering, University of Cambridge 2020
By downloading this dataset, you agree to the Creative Commons Attribution-NonCommercial 4.0 International license. This license allows users to use, share and adapt the dataset, so long as credit is given to the authors (e.g. by citation) and the dataset is not used for any commercial purposes.
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.