Jigsaw: Learning to Assemble Multiple Fractured Objects

Authors: Jiaxin Lu, Yifan Sun, Qixing Huang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Jigsaw on the Breaking Bad dataset and achieve superior performance compared to state-of-the-art methods.
Researcher Affiliation Academia Jiaxin Lu Yifan Sun Qixing Huang Department of Computer Science University of Texas at Austin {lujiaxin, yifansun12}@utexas.edu huangqx@cs.utexas.edu
Pseudocode No The paper describes the method components and pipeline in text and diagrams (Figure 1), but it does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://jiaxin-lu.github.io/Jigsaw/.
Open Datasets Yes We leverage the Breaking Bad dataset (Sellan et al., 2022), a novel data set of multiple fracture assemblies featuring synthetic physical breaking patterns.
Dataset Splits No Everyday consists of 498 models and 41,754 fracture patterns, and is split into a training set (34,075 fracture patterns from 407 objects) and a test set (7,679 fracture patterns from 91 objects). The paper specifies training and test splits but does not explicitly define a separate validation split.
Hardware Specification Yes All experiments are conducted on a Linux workstation with 4 Tesla V100-SXM2-32GB GPUs, Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz CPUs, and 480GB Memory.
Software Dependencies No The paper states 'Our framework is implemented in Pytorch' but does not provide specific version numbers for PyTorch or any other software dependencies used in the experiments.
Experiment Setup Yes Table 1 lists detailed experiment parameters including training epochs (250), batch size (4), learning rate (0.001), optimizer (Adam), scheduler (Cosine), minimum learning rate (1e-5), loss ratios (α, β, γ), and temperature parameter (τ). Additionally, it states: 'For Jigsaw, we start training with only the segmentation loss. We add matching loss after first 10 epochs, and rigidity loss after 200 epochs.'