Hard Example Generation by Texture Synthesis for Cross-domain Shape Similarity Learning
Authors: Huan Fu, Shunming Li, Rongfei Jia, Mingming Gong, Binqiang Zhao, Dacheng Tao
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach shows state-of-the-art performance on a recently released large-scale 3D-FUTURE [1] repository, as well as three widely studied benchmarks, including Pix3D [2], Stanford Cars [3], and Comp Cars [4]. Codes will be made publicly available at: https://github.com/3D-FRONT-FUTURE/IBSR-texture. We also conduct various ablation studies to comprehensively discuss our method. |
| Researcher Affiliation | Collaboration | Huan Fu1 Shunming Li1 Rongfei Jia1 Mingming Gong2 Binqiang Zhao1 Dacheng Tao3 1Tao Xi Technology Department, Alibaba Group, China 2The University of Melbourne, Australia 3The University of Sydney, Australia |
| Pseudocode | No | The paper describes its proposed methods and architecture but does not include any explicit pseudocode blocks or algorithms labeled as such. |
| Open Source Code | Yes | Codes will be made publicly available at: https://github.com/3D-FRONT-FUTURE/IBSR-texture. |
| Open Datasets | Yes | Our approach shows state-of-the-art performance on a recently released large-scale 3D-FUTURE [1] repository, as well as three widely studied benchmarks, including Pix3D [2], Stanford Cars [3], and Comp Cars [4]. |
| Dataset Splits | Yes | For Pix3D, We report both the category-specific scores and the mean scores as [15]. For 3D-FUTURE, we have 25,913 images and 4,662 3D shapes for our train set, and 5,865 images and 886 3D shapes for the validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models used for running the experiments. It mentions network architectures and batch sizes but no hardware specifications. |
| Software Dependencies | No | The paper mentions software components like 'Adam solver', 'U-Net', 'Patch GANs', 'Res Net-18', 'Res Net-34', and 'Blender [34]', but it does not specify version numbers for any of these or for any programming languages or deep learning frameworks used. |
| Experiment Setup | Yes | We first train TSM...We use the Adam solver [27] with a learning rate of 0.0002 and coefficients of (0.5, 0.999). The latent dimension |z| is set to 8...We train the network with a image size of 256 256 and with a batch size of 16. For the AMV-ML procedure, we train the network in two stages, i.e., a warm-up stage with Linst for 20 epochs and a fine-tuning stage with the objective LAMV ML for another 60 epochs. We use the SGD solver with a beginning learning rate of 0.002, the momentum of 0.9, and the weight decay of 0.0004. The learning rate is fixed in the initial 10 \ 20 epochs, and linearly decays to zero over the next 10 \ 40 epochs. The model is trained with a batch size of 24 and with a image size of 224 224. |