Higher-Order Function Networks for Learning Composable 3D Object Representations
Authors: Eric Mitchell, Selim Engin, Volkan Isler, Daniel D Lee
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We study the effectiveness of our method through various experiments on subsets of the Shape Net dataset. We find that the proposed approach can reconstruct encoded objects with accuracy equal to or exceeding state-of-the-art methods with orders of magnitude fewer parameters. |
| Researcher Affiliation | Collaboration | 1Stanford University 2Samsung AI Center New York 3University of Minnesota |
| Pseudocode | No | The paper describes the model architecture and procedures in text and diagrams but does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | See https://saic-ny.github.io/hof for code and additional information. |
| Open Datasets | Yes | We demonstrate the effectiveness of HOF on the task of 3D reconstruction from an RGB image using a subset of the Shape Net dataset (Chang et al., 2015). The dataset can be downloaded from https://github.com/xcyan/nips16_PTN. In our second experiment, we use a broader dataset based on Shape Net, with train and test splits taken from Tatarchenko et al. (2019). The dataset can be downloaded from https://github.com/lmb-freiburg/what3d. |
| Dataset Splits | Yes | The dataset contains 31773 ground truth point cloud models for training/validation and 7926 for testing. |
| Hardware Specification | Yes | All GPU experiments were performed on NVIDIA GTX 1080 Ti GPUs. The CPU running times were computed on one of 12 cores of an Intel 7920X processor. |
| Software Dependencies | No | The paper mentions software components like "Adam Optimizer" and "Re LU activation function" but does not specify their version numbers or the versions of broader frameworks/libraries (e.g., PyTorch, TensorFlow). |
| Experiment Setup | Yes | We use the Adam Optimizer with learning rate 1e-5 and batch size 1, training for 4 epochs for all experiments (1 epoch 725k parameter updates). |