Put the TotalCaputure_60FPS_Original folder inside data/source/ as well. The folder also packs Ground-Truth poses in old SMPL format which we do not use in this codebase - we simply use SMPL poses from the TotalCapture included in AMASS. Ĭontact the DIP authors ( ) to send you TC IMU signals preprocessed by them (in DIP format), the folder should carry the name “TotalCaputure_60FPS_Original”. You could skip this if you are not interested in evaluating your trained model on real IMUs from TC.)Ĭheck the licence terms and citation information for TotalCapture dataset. This folder should contain both IMU signals and SMPL ground-truth poses, both of which we use to train the model together with synthesized AMASS data. Put the DIP_IMU folder inside data/source/. Decompress the datasets and place these folders AMASS_CMU, KIT, Eyes_Japan_Dataset, HUMAN4D, ACCAD, DFaust_67, HumanEva, MPI_Limits, MPI_mosh, SFU, Transitions_mocap, TotalCapture, DanceDB inside data/source/ Real IMU signals from DIPĭownload DIP from. We only used a couple of subsets within AMASS to train the model including more data synthesized from AMASS might improve performance, but we have not tried it. SMPL+H is the file format we use, though no hand motions are generated by our algorithm. Decompress and replace the empty data/smplh folder from this repo with the decompressed folder. (Only tested on Ubuntu 18.04 Might work on Windows with some minor modifications)ġ.Go to ( ) and install the Python 3 version of Anaconda or Miniconda.Ģ.Open a new terminal and run the following commands to create a new conda environment ( ):ģ.Activate & enter the new environment you just creared:Ĥ.Inside the new environment, and inside the project directory:ĥ.Install pytorch with CUDA (only tested with the following version, should work with other versions though):Ĭonda install pytorch=1.7.1 cudatoolkit=10.2 -c pytorch (check pytorch website for your favorite version)Ħ.Install our fork of the Fairmotion library, at a location you prefer:ĭownload the SMPL model from. This is the Python implementation accompanying our TIP paper at SIGGRAPH Asia 2022.Ĭopyright 2022 Meta Inc. Transformer Inertial Poser (TIP): Real-time Human Motion Reconstruction from Sparse IMUs with Simultaneous Terrain Generation
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |