Deep Learning on Implicit Neural Representations of Shapes
Authors: Luca De Luigi, Adriano Cardace, Riccardo Spezialetti, Pierluigi Zama Ramirez, Samuele Salti, Luigi di Stefano
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we put forward this research problem and propose inr2vec, a framework that can compute a compact latent representation for an input INR in a single inference pass. We verify that inr2vec can embed effectively the 3D shapes represented by the input INRs and show how the produced embeddings can be fed into deep learning pipelines to solve several tasks by processing exclusively INRs. (Abstract) / In this section, we first present the set-up of our experiments. Then, we show how several tasks dealing with 3D shapes can be tackled by working only with inr2vec embeddings as input and/or output. Additional details on the architectures and on the experimental settings are in Appendix F. (Section 4, Deep Learning on INRs) |
| Researcher Affiliation | Academia | Luca De Luigi , Adriano Cardace , Riccardo Spezialetti University of Bologna {luca.deluigi4, adriano.cardace2@unibo.it, riccardo.spezialetti}@unibo.it Pierluigi Zama Ramirez, Samuele Salti, Luigi Di Stefano University of Bologna |
| Pseudocode | No | The paper describes algorithms in prose, such as in Appendix C detailing the procedure to sample a dense point cloud, but it does not present them in formal pseudocode blocks or algorithm listings. |
| Open Source Code | Yes | Additional details and code can be found at https://cvlab-unibo.github.io/inr2vec. (Section 1, Introduction) |
| Open Datasets | Yes | Point cloud retrieval... Model Net40 (Wu et al., 2015) and Shape Net10 (Chang et al., 2015)). (Section 4, Point cloud retrieval) / Shape classification... Scan Net10 (Dai et al., 2017). (Section 4, Shape classification) / Point cloud part segmentation... Shape Net Part Segmentation dataset (Yi et al., 2016)... (Section 4, Point cloud part segmentation) |
| Dataset Splits | Yes | Since only the train and test splits are released for all the datasets, we created validation splits from the training sets in order to follow a proper train/val protocol for both the baselines and inr2vec. (Section 4, Shape classification) / During training, we select the best model by evaluating its reconstruction capability on a validation set of INRs. (Appendix F, inr2vec training) |
| Hardware Specification | Yes | We implemented our framework with the Py Torch library, performing all the experiments on a single NVIDIA 3090 RTX GPU. (Appendix G, Implementation, Hardware and Timings) / All the times are computed on a gpu NVidia RTX 2080 Ti. (Table 10 footnote) |
| Software Dependencies | No | The paper mentions software like the 'Py Torch library' (Appendix G), 'Adam optimizer (Kingma & Ba, 2014)', 'Adam W optimizer (Loshchilov & Hutter, 2017)', 'Py Torch3D (Ravi et al., 2020)', and 'Open3D library (Zhou et al., 2018)'. However, it does not specify version numbers for PyTorch, PyTorch3D, or Open3D, which are necessary for full reproducibility. |
| Experiment Setup | Yes | In all the experiments reported in this section, we convert 3D discrete representations into INRs featuring 4 hidden layers with 512 nodes each... We train inr2vec using an encoder composed of four linear layers with respectively 512, 512, 1024 and 1024 features, embeddings with 1024 values and an implicit decoder with 5 hidden layers with 512 features. (Section 4, General settings) / We train inr2vec with Adam W optimizer (Loshchilov & Hutter, 2017), learning rate 1e-4 and weight decay 1e-2 for 300 epochs... processing at each training step a mini-batch of 16 INRs. (Appendix F, inr2vec training) |