Pose-Guided 3D Human Generation in Indoor Scene

Authors: Minseok Kim, Changwoo Kang, Jeongin Park, Kyungdon Joo

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our evaluations, we achieve the improvements with more plausible interactions and more variety of poses than prior research in qualitative and quantitative analysis. We evaluate the proposed 3D human generation framework in various aspects. Specifically, we first describe our implementation details including datasets in Sec. 4.1. We then quantitatively and qualitatively evaluate the proposed method in Sec. 4.2 and perform ablation study in Sec. 4.3.
Researcher Affiliation Academia Artificial Intelligence Graduate School, UNIST {hello96min, kangchangwoo, jeonginpark, kyungdon}@unist.ac.kr
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Project page: https://bupyeonghealer.github.io/phin/.
Open Datasets Yes We use the most widely used benchmarks in 3D human generation methods for evaluation: the PROX dataset (Hassan et al. 2019) and the PROX-E dataset (Zhang et al. 2020b). For the 3D human pose classification, we modify the PROX and PROX-E datasets in terms of human pose, and we call it PROX-Pose, in short the PROX-P dataset.
Dataset Splits No The paper specifies training and testing data but does not explicitly mention a validation set or its split details for reproduction.
Hardware Specification Yes All experiments are implemented in Pytorch v1.7.1 (Paszke et al. 2019) with Nvidia RTX 3090 GPU.
Software Dependencies Yes All experiments are implemented in Pytorch v1.7.1 (Paszke et al. 2019) with Nvidia RTX 3090 GPU.
Experiment Setup Yes We set {αdist, αnormal} = {0.002, 0.003} for the geometric alignment loss. For the generation loss, we set {αkl, αvp, αcoll, αrec} = {1, 0.001, 0.1, 0.001}, where αkl increases linearly in an annealing scheme (Bowman et al. 2015) for training. We use the Adam optimizer (Kingma and Ba 2014) with the learning rate 3e 4. The batch size is set to 32. Our model is trained 30 epochs, which takes around 1 day.