[Re] $p$-Poisson surface reconstruction in curl-free flow from point clouds
Abstract: This study presents a reproducibility analysis of the $p$-Poisson surface reconstruction method presented by Park et al. (NeurIPS 2023). The method utilizes the $p$-Poisson equation and a curl-free constraint for improved surface reconstruction from point clouds, claiming significant advancements over existing implicit neural representation techniques. This study evaluates the reproducibility and generalizability of the results reported in the original paper, focusing on the evaluation using the Surface Reconstruction Benchmark (SRB) dataset. The neural network architecture and training procedures are entirely re-implemented from scratch, emphasizing correctness and efficient execution. While the replication generally outperforms the four alternative methods mentioned in the original paper, the distance results reported in the original paper fail to be reproduced by the re-implementation. Notably, training with the code published in the original paper yields similar results to the reproduced results, still deviating from the findings presented in the original paper. The presented implementation demonstrates a significant improvement in training performance, achieving a five-fold acceleration in training times compared to the code used in the original paper by vectorizing the gradient calculations and leveraging just-in-time compilation of the training loop, which gives an actionable insight for others to explore and integrate such optimizations into their machine learning code. The re-implementation is available at \footnote{\url{https://anonymous.4open.science/r/pinc-B7CD}}.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Mathieu_Salzmann1
Submission Number: 2258
Loading