FNeVR: Neural Volume Rendering for Face Animation

Authors: Bohan Zeng, Boyu Liu, Hong Li, Xuhui Liu, Jianzhuang Liu, Dapeng Chen, Wei Peng, Baochang Zhang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our FNe VR obtains the best overall quality and performance on widely used talking-head benchmarks.
Researcher Affiliation Collaboration 1Institute of Artificial Intelligence, School of Automation Science and Electrical Engineering, Beihang University, Beijing, P.R.China 2Sino-French Engineer School, Beihang University, Beijing, P.R.China 3Huawei Noah s Ark Lab, Shenzhen, P.R.China 4Huawei, Shenzhen, P.R.China 5Zhongguancun Laboratory, Beijing, P.R.China
Pseudocode No The paper includes architectural diagrams and descriptions of methods but no explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available1. 1https://github.com/zengbohan0217/FNe VR
Open Datasets Yes Our evaluation is performed on Vox Celeb [34], which contains more than 100,000 videos covering 1,251 speakers of different identities, and Vox Celeb2 [9], which contains about 1M videos of different celebrities.
Dataset Splits No The paper mentions training and testing sets, and evaluation on 'Vox Celeb' and 'Vox Celeb2', but does not provide explicit percentages or sizes for training/validation/test splits, nor does it explicitly mention a distinct validation set with quantitative details.
Hardware Specification Yes Since our model is lightweight, we only use 2 24GB NVIDIA 3090 GPUs during training.
Software Dependencies No The paper mentions 'Mind Spore' and 'CANN' but does not specify their version numbers or other software dependencies with version details.
Experiment Setup Yes FNe VR is trained for 100 epochs, repeating the video image set 75 times per epoch. We adopt Adam [29] optimizer with learning rate η = 2 10 4, β1 = 0.5 and β2 = 0.9 for each module.