Neural Processing of Tri-Plane Hybrid Neural Fields
Authors: Adriano Cardace, Pierluigi Zama Ramirez, Francesco Ballerini, Allan Zhou, Samuele Salti, Luigi di Stefano
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we show that the tri-plane discrete data structure encodes rich information, which can be effectively processed by standard deep-learning machinery. We define an extensive benchmark covering a diverse set of fields such as occupancy, signed/unsigned distance, and, for the first time, radiance fields. While processing a field with the same reconstruction quality, we achieve task performance far superior to frameworks that process large MLPs and, for the first time, almost on par with architectures handling explicit representations. To validate our results, we build a comprehensive benchmark for tri-plane neural field classification. We perform extensive tests to validate our approach. |
| Researcher Affiliation | Academia | Adriano Cardace1 Pierluigi Zama Ramirez1 Francesco Ballerini1 Allan Zhou2 Samuele Salti1 Luigi Di Stefano1 1 University of Bologna 2 Stanford University adriano.cardace2@unibo.it |
| Pseudocode | No | No structured pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | Yes | Summary of our contributions. Code available at https://github.com/CVLAB-Unibo/ triplane_processing. |
| Open Datasets | Yes | Specifically, we test all methods on UDF fields obtained from point clouds of Model Net40 (Wu et al., 2015), Shape Net10 (Qin et al., 2019), and Scan Net10 (Qin et al., 2019); SDF fields learned from meshes of Manifold40 (Hu et al., 2022); OF fields obtained from voxels grids of Shape Net10. In addition, we provide for the first time classification results on neural radiance fields (RF), learned from Shapenet Render (Xu et al., 2019). We additionally note that the Shape Net10 and Scan Net10 datasets mentioned in the main text are subsets of 10 (shared) classes of Shape Net (Chang et al., 2015) and Scan Net (Dai et al., 2017), respectively, originally proposed in Qin et al. (2019). |
| Dataset Splits | Yes | For the train/val/test splits, we randomly sample 80% of the objects for the training set and select 10% shapes for both the validation and test splits. Accordingly, we build training, validation, and test sets with samples drawn from all three fields. The inr2vec framework (De Luigi et al., 2023) was trained on each point cloud, mesh, and voxel dataset for 100 epochs via the official code and the resulting embeddings used to train a classifier (3 fully connected layers with Re LUs). |
| Hardware Specification | Yes | We use a single NVIDIA RTX 3090 for all experiments. Time (hours) ... Times are computed on a single NVIDIA RTX 3090. |
| Software Dependencies | No | No specific version numbers for key software dependencies like Python, PyTorch, or CUDA are provided. It mentions "Nerf Acc library" but without a version. |
| Experiment Setup | Yes | We adopt the Adam W optimizer (Loshchilov & Hutter, 2017), with the One Cycle scheduler (Smith & Topin, 2019) with maximum learning rate of 1e-4, and train for 150 epochs with batch size set to 256 using the cross entropy loss. As for part segmentation, we use the same configuration as for classification, although we train for 75 epochs with batch size 32. |