Learning Adversarial 3D Model Generation With 2D Image Enhancer
Authors: Jing Zhu, Jin Xie, Yi Fang
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The evaluations on two large-scale 3D model datasets, Shape Net and Model Net, demonstrate that our proposed method can not only generate highquality 3D models, but also successfully learn discriminative shape representation for classification and retrieval without supervision.To comprehensively validate the performance of our proposed method, we conduct experiments on two large-scale datasets for different tasks, including 3D model generation, shape classification and shape retrieval. |
| Researcher Affiliation | Academia | Jing Zhu, Jin Xie, Yi Fang NYU Multimedia and Visual Computing Lab Department of Electrical and Computer Engineering, NYU Abu Dhabi, UAE Department of Electrical and Computer Engineering, NYU Tandon School of Engineering, USA Department of Computer Science and Engineering, NYU Tandon School of Engineering, USA |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not include an unambiguous statement or link for the open-source code of the methodology described. It states, "Since the author of 3D-GAN did not provide a pretrained model for feature extraction or source code of a trainable model, we cannot obtain the retrieval performance using the original 3D-GAN model.", but this refers to another work. |
| Open Datasets | Yes | In this task, we train our proposed framework on large-scale Shape Net (Chang et al. 2015) dataset that contains more than 50, 000 3D models with 55 categories. 3DR2N2 (Choy et al. 2016) provides a dataset that includes rendered images from 3D models in Shape Net from 23 different views. Then, we input 3D models from Model Net (Wu et al. 2015) into the trained 3D model discriminator and concatenate the features extracted from each convolution layer (after max pooling) as shape representations. |
| Dataset Splits | Yes | The Model Net10 subset contains 4, 899 models from 10 different categories, which are split into a training set with 3, 991 models and a testing set with 908 models. The Model Net40 subset has a total of 12, 311 models from 40 categories, split into a training set and a testing set with size 9, 843 and 2, 468, respectively. |
| Hardware Specification | Yes | We implement our framework using the popular deep learning tool Tensor Flow (Abadi et al. 2016) and train it on a desktop with Intel Xeon E5-2603 CPU and NVIDIA Tesla K80 GPU. |
| Software Dependencies | No | The paper mentions "Tensor Flow (Abadi et al. 2016)", but it does not specify a version number for TensorFlow or any other software libraries or dependencies used in the experiments. |
| Experiment Setup | Yes | We use ADAM (Kingma and Ba 2014) optimizer to obtain the optimal network parameters with beta value β = 0.5 and learning rate 0.0002 for generators and discriminators. The batch size is set to 64. |