Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Point Cloud Synthesis Using Inner Product Transforms
Authors: Ernst Röell, Bastian Rieck
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness, efficiency, and overall utility of our IP-Encoder through a comprehensive suite of experiments. Our first set of experiments assesses (i) the reconstruction quality of the IP-Encoder, (ii) the efficacy of the IPT as a loss, and (iii) its computational efficiency. Our IP-Encoder consistently ranks second in terms of the EMD, known to be the most suitable metric to evaluate reconstruction quality [44]. We additionally perform an ablation study concerning the loss term, showing that the combination of CD and IPT is crucial for high reconstruction quality. As Table 2 shows, a joint loss yields the best quality. |
| Researcher Affiliation | Academia | 1AIDOS Lab, University of Fribourg, Switzerland 2Institute of AI for Health, Helmholtz Munich, Germany 3Technical University of Munich, Germany |
| Pseudocode | No | The paper describes the methodology in prose and mathematical formulations. There are no explicitly labeled pseudocode blocks or algorithms presented in a structured, code-like format. |
| Open Source Code | Yes | Our code is available at https://github.com/aidos-lab/inner-product-transforms and is released under a BSD-3-Clause license. |
| Open Datasets | Yes | All our experiments rely on a subset of the Shape Net dataset, and we adopt the preprocessing and evaluation workflow introduced by Yang et al. [40]. We train the IP-Encoder and IP-VAE on the MNIST dataset. |
| Dataset Splits | No | Each point cloud in the dataset consists of 2048 points sampled on the surface of three shape classes (airplanes, chairs and cars). Our evaluation of generative performance follows the setup of Yang et al. [40]. The paper does not explicitly state the train/test/validation splits used for its experiments, only referring to the setup of a cited work without detailing the specific splits within this paper. |
| Hardware Specification | Yes | Our hardware consists of an NVIDIA GeForce RTX 4070 with 12GB VRAM and an 13th Gen Intel(R) Core(TM) i7-13700K with 32GB RAM. |
| Software Dependencies | No | The paper mentions various architectural components (e.g., CNN, VAE) and optimizers (Adam) but does not provide specific version numbers for any software libraries or frameworks used in its implementation. |
| Experiment Setup | Yes | We consider the IPT as a 1D signal, with each direction corresponding to a channel, sample 128 directions and discretise each direction into 128 steps, thus obtaining an IPT with a resolution of 128 128. Our IP-Encoder consists of four 1D convolutional layers with batch normalisation, max-pooling, and SiLU activation functions (cf. Appendix F). We train the IP-Encoder separately for each of the classes for 5k epochs, using a CD + IPT-64 loss. For training the IP-VAE, we follow the β-VAE setup, using KL-divergence and MSE loss terms with β = 1.00 10 4. For this experiment, we apply backpropagation for 2000 epochs with the Adam optimiser using a learning rate of 0.5, which is halved in epochs 50, 100, 200, and 1000, respectively. |