Point Cloud Completion with Pretrained Text-to-Image Diffusion Models

Authors: Yoni Kasten, Ohad Rahamim, Gal Chechik

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate SDS-Complete on a collection of incomplete scanned objects, captured by real-world depth sensors and Li DAR scanners. We find that it effectively reconstructs objects that are absent from common datasets, reducing Chamfer loss by about 50% on average compared with current methods.
Researcher Affiliation Collaboration Yoni Kasten1 Ohad Rahamim2 Gal Chechik1,2 1NVIDIA Research 2Bar-Ilan University
Pseudocode No No clearly labeled pseudocode or algorithm block was found in the paper.
Open Source Code Yes Project page: https://sds-complete.github.io/
Open Datasets Yes For depth images, we used the Redwood dataset [12] that contains a diverse set of objects. ... We further tested our model on the KITTI Li DAR dataset [7, 17], which contains incomplete point clouds of objects in real-world scenes captured by Li DAR sensors.
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset split percentages or sample counts for their experiments.
Hardware Specification Yes All times are measured when running our method on NVIDIA RTX A6000.
Software Dependencies Yes As a text-to-image diffusion model we use Stable Diffusion v2 [43].
Experiment Setup Yes We optimize the networks using the Adam optimizer [26] with a learning rate 10 4. The coefficients for our loss for all the experiments are δm = 105, δd = 105, δp = 105, δeikonal = 104, δplane = 105. At each iteration we sample 1000 uniform points for Lplane and Leikonal.