Shape-Pose Ambiguity in Learning 3D Reconstruction from Images

Authors: Yunjie Wu, Zhengxing Sun, Youcheng Song, Yunhan Sun, YiJie Zhong, Jinlong Shi2978-2985

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments on synthetic and real image datasets, we demonstrate that our method can perform comparably to existing methods while not requiring any extra pose-aware annotations, making it more applicable and adaptable. We conduct comprehensive experiments on public datasets, including both synthetic and real images.
Researcher Affiliation Academia Yunjie Wu,1 Zhengxing Sun,1 Youcheng Song,1 Yunhan Sun,1 Yi Jie Zhong,1 Jinlong Shi2 1 State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, P R China 2 Department of Computer, Jiangsu University of Science and Technology, Zhengjiang, P R China szx@nju.edu.cn
Pseudocode No The paper describes its method in detail using text and diagrams, but it does not include any pseudocode or algorithm blocks.
Open Source Code Yes We provide more details in the appendix and the code in: https://github.com/Jiejiang Wu/Shape-Pose-Ambiguity.
Open Datasets Yes Datasets We test our method on two public datasets. The first is the Shape Net s rendered images (Kato, Ushiku, and Harada 2018). ... The second is CUB-200-2011 (Wah et al. 2011). It contains images of 200 species of birds.
Dataset Splits Yes We use the provided train-test split. For both two datasets, we use no extra annotations except the silhouettes for training. ... We use an 8:2 train-test split.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only mentions general settings like 'The batch size is 32'.
Software Dependencies No The paper mentions using a 'differentiable renderer (Liu et al. 2019)' and 'resnet-18 (He et al. 2016)', but it does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Each shape has 642 vertices, the same as VPL (Kato and Harada 2019). ... The batch size is 32, and the learning rate is 1e-4. We provide more details in the appendix and the code in: https://github.com/Jiejiang Wu/Shape-Pose-Ambiguity. ... In practice, the λ is set to 5e-4, and the μ is set to 5e-3 to make all loss terms in the close magnitude.