MV2Cyl: Reconstructing 3D Extrusion Cylinders from Multi-View Images

Authors: Eunji Hong, Minh Hieu Nguyen, Mikaela Angelina Uy, Minhyuk Sung

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments, comparing our method with previous work that takes a raw 3D point cloud as input, demonstrate the effectiveness of our approach by taking advantage of multi-view images. [...] Tab. 1 shows the quantitative results of CAD reconstruction on the Fusion360 [64] and Deep CAD [65] datasets. We see that MV2Cyl outperforms the baselines by considerable margins across both datasets in all evaluated metrics.
Researcher Affiliation Collaboration Eunji Hong 1, Minh Hieu Nguyen1, Mikaela Angelina Uy2,3, Minhyuk Sung1 1Korea Advanced Institute of Science and Technology, 2Stanford University, 3NVIDIA
Pseudocode No The paper describes its method through text and figures (e.g., Figure 1, Figure 4 showing pipelines and steps), but it does not contain any structured pseudocode or algorithm blocks with numbered or bulleted code-like steps explicitly labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code No The paper states 'Our project page can be found at https://mv2cyl.github.io.' However, in the NeurIPS Paper Checklist, the authors state for question '5. Open access to data and code': 'Answer: [No] Justification: The code and data will be made available upon the acceptance of this work.' This indicates that the code is not currently openly accessible.
Open Datasets Yes Datasets. We evaluate our approach on two sketch-extrude CAD datasets Fusion360 [64] and Deep CAD [65]. We use the train and test splits as released in [58].
Dataset Splits Yes The datasets are divided into training and validation sets in a 9:1 ratio for the U-Net [50] training.
Hardware Specification Yes The training times for 3 epochs are 316 minutes for the Fusion360 [64] dataset and 381 minutes for the Deep CAD dataset, using a single NVIDIA RTX A6000 GPU. [...] We trained all the fields for 1500 iterations on a single NVIDIA RTX 3090 GPU.
Software Dependencies No The paper mentions 'Blender Proc [7]' and 'Sci Py library [59]' but does not provide specific version numbers for these software components or any other key libraries or frameworks used in the implementation.
Experiment Setup Yes We fix K = 8 instances following [58] and train with image resolutions of H = W = 400. [...] The training objective is defined as: ext{L}_{ ext{2D} ext{curve}}= ext{lambda}_{ ext{CE}} ext{L}_{ ext{cross-entropy}} + ext{lambda}_{ ext{focal}} ext{L}_{ ext{focal}} + ext{lambda}_{ ext{dice}} ext{L}_{ ext{dice}}, where ext{lambda}_{ ext{CE}}, ext{lambda}_{ ext{focal}}, and ext{lambda}_{ ext{dice}} are all set to 1.0... The opacity field ext{sigma}( ext{x}) hyper-parameters are set to ext{zeta}=10 and ext{beta}=0.8 as in [68]. For the reconstruction loss of the density field ext{F}, we set ext{lambda}_{ ext{sparsity}}=0.5, s = 0.5, ext{eta}_{ ext{batch}}=0.1, and ext{eta}_{ ext{image}}=0.05. We use a batch size of 8,192 rays for the training of both the density field ext{F} and the semantic field ext{A}. We trained all the fields for 1500 iterations...