Fully Convolutional Mesh Autoencoder using Efficient Spatially Varying Kernels
Authors: Yi Zhou, Chenglei Wu, Zimo Li, Chen Cao, Yuting Ye, Jason Saragih, Hao Li, Yaser Sheikh
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, We first examine the generality of our AE models on different types of 3D mesh data. Then, we present localized latent code interpolation for 3D hand models. After that, we compare our model with SOTA 3D mesh AEs. Finally we compare the performance of different convolution and (un)pooling layers under the same network architecture. All experiments were trained with L1 reconstruction loss only, Adam [17] optimizer and reported with point to point mean euclidean distance error if not specified. Additional experiment details can be found in the appendix. |
| Researcher Affiliation | Collaboration | Yi Zhou Adobe Research Chenglei Wu Facebook Reality Labs Zimo Li University of Southern California Chen Cao Facebook Reality Labs Yuting Ye Facebook Reality Labs Jason Saragih Facebook Reality Labs Hao Li Pinscreen Yaser Sheikh Facebook Reality Labs |
| Pseudocode | No | The paper describes the operations and their mathematical formulations but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain an explicit statement about open-sourcing the code or a link to a code repository. |
| Open Datasets | Yes | We first experimented on the 2D-manifold D-FAUST human body dataset [4]. It contains 140 sequences of registered human body meshes. We used 103 sequences (32933 meshes in total) for training, 13 sequences (4023 meshes) for validation and 13 sequences (4264 meshes) for testing. |
| Dataset Splits | Yes | We used 103 sequences (32933 meshes in total) for training, 13 sequences (4023 meshes) for validation and 13 sequences (4264 meshes) for testing. |
| Hardware Specification | No | The paper mentions 'GPU memory consumption' but does not specify any particular GPU or CPU models, or other hardware details used for the experiments. |
| Software Dependencies | No | The paper mentions 'Adam [17] optimizer' and 'Py Torch Geometry [13]' but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | All experiments were trained with L1 reconstruction loss only, Adam [17] optimizer and reported with point to point mean euclidean distance error if not specified. Additional experiment details can be found in the appendix. Our AE has a mean test error at 5.01 mm after 300 epochs of training. After 400 epochs of training, the error converged to 0.2 mm. After 36 epochs, the reconstruction error dropped to 4.1 cm. After training 100 epochs, the mean test error dropped to 3.01 mm. Both networks are set with compression ratio (network bottle neck size over original size) at 0.3%. We trained ours with 200 epochs and Neural3DMM with 300 epochs. Table 1 lists all the block designs, the errors, the parameter count and the GPU memory consumption for training with batch size=16. |