Banana: Banach Fixed-Point Network for Pointcloud Segmentation with Inter-Part Equivariance
Authors: Congyue Deng, Jiahui Lei, William B Shen, Kostas Daniilidis, Leonidas J. Guibas
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through experiments conducted on both articulated objects and multi-object scans, we demonstrate the efficacy of our approach in achieving strong generalization under inter-part transformations, even when confronted with substantial changes in pointcloud geometry and topology. and 5 Experiments |
| Researcher Affiliation | Academia | Congyue Deng1 Jiahui Lei2 Bokui Shen1 Kostas Daniilidis2,3 Leonidas Guibas1 1 Stanford University 2 University of Pennsylvania 3 Archimedes, Athena RC {congyue, willshen, guibas}@cs.stanford.edu, {leijh, kostas}@cis.upenn.edu |
| Pseudocode | No | The paper describes the network architecture and message passing verbally and with diagrams (Figure 3), but does not provide a formally labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | We begin by evaluating our method on articulated-object part segmentation using on four categories of the Shape2Motion dataset [74]: washing machine, oven, eyeglasses, and refrigerator. and Dyn Lab [30] is a collection of scanned laboratory scenes, each with 2-3 rigidly moving solid objects captured under 8 different configurations with object positions randomly changed. |
| Dataset Splits | No | The paper describes training on 'single rest state' and testing on 'articulated states' for Shape2Motion, and overfitting to the 'first configuration' for Dyn Lab, but does not provide specific train/validation/test dataset splits (e.g., percentages, sample counts, or citations to predefined splits with such details). |
| Hardware Specification | No | The paper does not provide specific details regarding the hardware used for running experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions using a 'SE(3)-equivariant framework called Vector Neurons (VN) [15]' but does not provide specific version numbers for any software dependencies or libraries used in the implementation or experiments. |
| Experiment Setup | Yes | To demonstrate the generalizability of our approach, we train our model on objects in a single rest state, such as ovens with closed doors (as depicted in Fig. 5, left). This training setup aligns with many synthetic datasets featuring static shapes [9, 52]. and For weight truncation [72], we use the l∞-Lipschitz as it is the loosest constraint and set the per-layer Lipschitz upper bound to be 0.99999. For regularization losses, we set the overall network Lipschitz threshold to 0.99. The adversarial sampling [69] loss only works with l2-norm. We reduce the total number of points from 2048 to 1024 and the GNN neighbors from 40 to 20 due to the large memory consumption for the loss computations. |