This repo contains instructions for running "PPR: Physically Plausible Reconstruction from Monocular Videos (ICCV 23)", which builds 4D models of the object and the environment by coupling differentiable rendering and differentiable physics simulation. For differentiable physics simulation (4D to physics), please refer to ppr-diffphysics. For differentiable rendering (video to 4D), please see lab4d.
Attention: Compared to lab4d main branch, ppr additionally extracts surface normal. Make sure you rerun bash scripts/install-deps.sh and the pre-processing script if the data was previously processed by lab4d main.
Follow instructions to install lab4d. Inside lab4d root directory, checkout the ppr branch and update the diffphys submodule.
git checkout ppr
git submodule update --recursive
Activate lab4d conda environment and install additional dependencie:
pip install -r projects/ppr/ppr-diffphys/requirements.txt
pip install urdfpy==0.0.22 --no-deps
pip install open3d==0.17.0
pip install geomloss==0.2.6
We show how PPR works using the cat-pikachu-0 sequence. Begin by downloading the pre-processed video data.
bash scripts/download_unzip.sh "https://www.dropbox.com/scl/fi/j2bztcv49mqc4a6ngj8jp/cat-pikachu-0.zip?rlkey=g2sw8te19bsirr5srdl17kzt1&dl=0"
Note: Processing follows the pre-processing pipeline of lab4d, except frames are extracted with a constant rate of 10 fps (no flow-guided filtering).
Training has 2 stages. We first optimize background and object. The following can be executed in parallel. To reconstruct the background scene:
# Args: gpu-id, sequence name, hyper-parameters defined in lab4d/config.py
bash scripts/train.sh lab4d/train.py 0 --seqname cat-pikachu-0 --logname bg --field_type bg --data_prefix full --num_rounds 60 --alter_flow --mask_wt 0.01 --normal_wt 1e-2 --reg_eikonal_wt 0.01
To reconstruct the object:
# Args: gpu-id, sequence name, hyper-parameters defined in lab4d/config.py
bash scripts/train.sh lab4d/train.py 1 --seqname cat-pikachu-0 --logname fg-urdf --fg_motion urdf-quad --num_rounds 20 --feature_type cse
The second stage is a physics-informed optimization that couples the object and the scene:
# Args: gpu-id sequence name, hyper-parameters in both lab4d/config.py and and projects/ppr/config.py
bash scripts/train.sh projects/ppr/train.py 0 --seqname cat-pikachu-0 --logname ppr --field_type comp --fg_motion urdf-quad --feature_type cse --num_rounds 20 --learning_rate 1e-4 --pixels_per_image 12 --iters_per_round 100 --secs_per_wdw 2.4 --noreset_steps --noabsorb_base --load_path logdir/cat-pikachu-0-fg-urdf/ckpt_latest.pth --load_path_bg logdir/cat-pikachu-0-bg/ckpt_latest.pth
You may find the physics simulation results at logdir/cat-pikachu-0-ppr/all-phys-00-05d.mp4.
Visualization at 0 iteration. Left to right: target, simulated, control reference, "distilled" simulation (to regularize diff-rendering).
all-kinematics-00-00000.mp4
Visualization at 1000 iteration:
all-phys-00-01000.mp4
Note: Multi-gpu physics-informed optimization is not supported as of now.
To visualize the intermediate simulated trajectories over the course of optimization:
python projects/ppr/render_intermediate.py --testdir logdir/cat-pikachu-0-ppr/ --data_class sim
sim.mp4
To export meshes and visualize results, run
python projects/ppr/export.py --flagfile=logdir/cat-pikachu-0-ppr/opts.log --load_suffix latest --inst_id 0 --vis_thresh -10 --extend_aabb
render-bone-compose-ref.mp4
There are a few other modes supported by the mesh renderer, such as bird's eye view, and ghosting.
python lab4d/render_mesh.py --testdir logdir/cat-pikachu-0-ppr/export_0000/ --view bev --ghosting
render-bone-compose-bev.mp4
We use viser for interactive visualization of the 4D scene.
Install viser by pip install viser. Once the result meshes have been exported to logdir/$logname/export_$inst_id, run
python lab4d/mesh_viewer.py --testdir logdir/cat-pikachu-0-ppr/export_0000/
gui.mp4
To run physics simulation with the optimized parameters (control reference, body mass, PD gains):
python projects/ppr/simulate.py --flagfile=logdir/cat-pikachu-0-ppr/opts.log --load_suffix latest --load_suffix_phys latest --inst_id 0
To visualize the output in the interactive mode:
python lab4d/mesh_viewer.py --testdir logdir//cat-pikachu-0-ppr-exp2/simulate_0000/sim_traj/
Oct-09-2023.21-10-12.mp4
Run the following script to get results on AMA samba and bouncing sequences.
bash projects/ppr/run_ama.sh
It downloads the pre-processed data and then runs training and visualization routines. We suggest reading the comments in the script to understand what happens in each step. Expected results:
render-bone-compose-bev.mp4
render-bone-compose-bev2.mp4
First download the AMA data and place it at database/ama/T_samba/{meshes/, calibration/}.
For running evaluation scripts, a few extra dependencies need to be installed:
pip install git+https://github.com/facebookresearch/pytorch3d
Once the result meshes have been exported to logdir/$logname/export_$inst_id, execute
python projects/ppr/eval/compute_metrics.py --testdir logdir//ama-bouncing-4v-ppr-exp/export_0000/ --gt_seq D_bouncing-1 --pred_prefix "" --fps 30
This computes chamfer distance and f-score on the mesh sequence by comparing it against the ground truth.
Results of DiffRen optimization-only:
Avg chamfer dist: 10.90cm
Avg f-score at d=10cm: 85.6%
Avg f-score at d=5cm: 61.4%
Avg f-score at d=2cm: 25.3%
Results of DiffRen+DiffSim optimization:
Avg chamfer dist: 8.85cm
Avg f-score at d=10cm: 90.6%
Avg f-score at d=5cm: 70.0%
Avg f-score at d=2cm: 32.3%
Note: The paper uses the old codebase so the the numbers might differ. We plan to revise the paper so it matches this codebase.