Convolutional Generation of Textured 3D Meshes

Authors: Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, Aurelien Lucchi

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of our method on Pascal3D+ Cars and CUB, both in an unconditional setting and in settings where the model is conditioned on class labels, attributes, and text.
Researcher Affiliation Academia Dario Pavllo Dept. of Computer Science ETH Zurich Graham Spinks Dept. of Computer Science KU Leuven Thomas Hofmann Dept. of Computer Science ETH Zurich Marie-Francine Moens Dept. of Computer Science KU Leuven Aurelien Lucchi Dept. of Computer Science ETH Zurich
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes We release our code and pretrained models at https://github.com/dariopavllo/convmesh.
Open Datasets Yes We evaluate our method on two datasets with annotated keypoints, and use the implementation of [24] to estimate the pose from keypoints using structure-from-motion. CUB-200-2011 [52] We use the train/test split of [24]... Pascal3D+ (P3D) [57] We use the cars subset...
Dataset Splits No The paper mentions "train/test split" for CUB and "training images" for P3D, but does not explicitly detail a validation set or its split.
Hardware Specification Yes We train with a batch size of 50 on a single Pascal GPU, which requires 12 hours.
Software Dependencies No The paper mentions software like Adam, DIB-R, and Mask R-CNN, but does not provide specific version numbers for them.
Experiment Setup Yes The model (Fig. 1) is trained for 1000 epochs using Adam [28], with an initial learning rate of 10 4 halved every 250 epochs. We train with a batch size of 50 on a single Pascal GPU, which requires 12 hours. (...) train for 600 epochs with a constant learning rate of 0.0001 for G and 0.0004 for D (two time-scale update rule [19]). We update D twice per G update, and evaluate the model on a running average of G s weights (β = 0.999) as proposed by [64, 25, 26, 4]. (...) For all experiments, we use a total batch size of 32 and we employ synchronized batch normalization across multiple GPUs.