LCollision: Fast Generation of Collision-Free Human Poses using Learned Non-Penetration Constraints

Authors: Qingyang Tan, Zherong Pan, Dinesh Manocha3913-3921

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have evaluated our method on the SCAPE dataset (Anguelov et al. 2005), the MIT-Swing dataset (Vlasic et al. 2008), and the MIT Jumping dataset (Vlasic et al. 2008). Combining these techniques, we achieve an accuracy of 94.1%, a false positive rate of 6.1%, and a false negative rate of 5.7% when predicting collisions for 2.5 106 randomized human poses sampled from these datasets. Moreover, our learned collision detector is 80 faster than prior exact collision detection methods running on a CPU (Pan, Chitta, and Manocha 2012). In Table 1, we compare the accuracy of baselines in terms of predicting penetration depth energies, ranking penetration depth energies, and classifying collision-free meshes.
Researcher Affiliation Academia Qingyang Tan1, Zherong Pan2, Dinesh Manocha1 1 Department of Computer Science, University of Maryland at College Park 2 Department of Computer Science, University of Illinois at Urbana-Champaign qytan@umd.edu, zherong@illinois.edu, dmanocha@umd.edu
Pseudocode Yes Algorithm 1 Generating Penetration Energy Vector PDe
Open Source Code No The paper does not provide any statement about releasing source code or a link to a code repository.
Open Datasets Yes We have evaluated our method on the SCAPE dataset (Anguelov et al. 2005), the MIT-Swing dataset (Vlasic et al. 2008), and the MIT Jumping dataset (Vlasic et al. 2008).
Dataset Splits Yes During the second stage, we use 0.7M samples of the augmented dataset for training and 0.3M samples for validation.
Hardware Specification Yes All the training and testing are performed on a single desktop machine with a 4-core CPU, 32GB memory, and an NVIDIA GTX 1080Ti GPU.
Software Dependencies No The paper mentions 'We implement our method using Py Torch (Paszke et al. 2017)', but does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes This stage uses the loss: L = w P DLP D + wrank Lrank + wentropy Lentropy, which is configured with w P D = 5, wrank = 2, wentropy = 2, and trained using a learning rate of 0.001 and a batch size of 32 over 30 epochs. We set Z0 = 10 for SCAPE and Z0 = 12 for Swing and Jumping.