Point Cloud Semantic Scene Completion from RGB-D Images

Authors: Shoulong Zhang, Shuai Li, Aimin Hao, Hong Qin3385-3393

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments validate that our new method achieves the stateof-the-art performance, in contrast with the current methods applied to our dataset.
Researcher Affiliation Collaboration 1Beihang University, Beijing, China 2Peng Cheng Laboratory, Shenzhen, China 3Stony Brook University (SUNY), Stony Brook, USA
Pseudocode Yes Algorithm 1: Main Steps of PCSSC-Net Training.
Open Source Code No The paper does not provide a direct link or explicit statement about releasing the source code for the described methodology.
Open Datasets Yes we synthesized a new dataset with 12 layouts of the Scene Net RGB-D dataset (Mc Cormac et al. 2017) and 263 typical indoor object models of the Shape Net dataset (Chang et al. 2015) with realistic textures.
Dataset Splits No We divide our dataset into 1520 and 392 scenes for the training and testing purpose. A separate validation split is not explicitly defined.
Hardware Specification Yes We trained our model on Nvidia RTX 2080Ti GPU for roughly 70 hours with a batch size of 8.
Software Dependencies No The paper states 'Our model is implemented in Py Torch' but does not specify its version number or versions for any other software dependencies.
Experiment Setup Yes The initial learning rate is set to 0.001, and the decay rate is 0.7 for every ten epochs. For the parameters in the loss, we set α = 0.005, β = 1.5 and γ = 0 for the first 10 training hours. Then, we set α = 0.015, β = 1.5, and γ = 1.5 for geometry refinement based on their semantic classes. The CD weight for the chair class is set to 0.01 and 1 for other classes.