A Simple yet Universal Framework for Depth Completion

Authors: Jin-Hwi Park, Hae-Gon Jeon

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the proposed method s superior generalization capabilities for Uni DC problem over state-of-the-art depth completion.
Researcher Affiliation Academia Jin-Hwi Park AI Graduate School GIST jinhwipark@gm.gist.ac.kr Hae-Gon Jeon AI Graduate School GIST haegonj@gist.ac.kr
Pseudocode Yes Algorithm 1 Implementation of Hyperbolic Universal Depth Completion
Open Source Code Yes Source code is publicly available at https://github.com/Jinhwi Park/Uni DC.
Open Datasets Yes We employ the NYU Depth V2 dataset, which consists of 464 indoor scenes captured using a Kinect sensor... For outdoor environments, we utilize the KITTI DC dataset, which comprises 90K samples.
Dataset Splits Yes The dataset is segmented into training (86K samples), validation (7K samples), and testing (1K samples) portions.
Hardware Specification Yes Our model is implemented with public Py Torch [92], trained on a single RTX 3090Ti GPU using Adam [93] optimizer.
Software Dependencies No The paper mentions 'public Py Torch [92]' and 'Adam [93] optimizer' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We train our method in a supervised manner with a linear combination of two loss terms... where µ is a balance term and is empirically set to 0.1. ... The initial learning rate was set to 5 × 10^−3 and reduced by 0.1 every 20% for total iterations.