SfPUEL: Shape from Polarization under Unknown Environment Light

Authors: Youwei Lyu, Heng Guo, Kailong Zhang, Si Li, Boxin Shi

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on synthetic and our collected real-world dataset demonstrate that Sf PUEL significantly outperforms existing Sf P and single-shot normal estimation methods.
Researcher Affiliation Collaboration Youwei Lyu1 Heng Guo1 Kailong Zhang1 Si Li1 Boxin Shi2,3 1School of Artificial Intelligence, Beijing University of Posts and Telecommunications 2State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University 3National Engineering Research Center of Visual Technology, School of Computer Science, Peking University
Pseudocode No The paper describes the network architecture and its components in text and diagrams, but does not provide pseudocode or algorithm blocks.
Open Source Code Yes The code and dataset is available at https://github.com/Youwei Lyu/Sf PUEL.
Open Datasets Yes To validate the effectiveness of our method under different material types, we further collect a real-world dataset as a complement to existing Sf P datasets (15; 4). ... We create a synthetic dataset by Mitsuba2 (41) to enable model training and validation. ... The code and dataset is available at https://github.com/Youwei Lyu/Sf PUEL.
Dataset Splits Yes We create a synthetic dataset by Mitsuba2 (41) to enable model training and validation. We collect 2,074 high-quality SVBRDFs, 799 HDR environment maps (1; 43), and 244 object meshes (1; 43; 50; 14) for data synthesis. We use 1,983 SVBRDFs, 651 environment maps, and 200 object meshes as the source data to generate the training dataset composed of 20,000 sets of images with the resolution of 512 512, and the remaining materials are adopted for rendering 1,000 validation data with the same resolution as training ones.
Hardware Specification Yes All experiments are conducted on Ubuntu 20.04 LTS with four NVIDIA RTX 3090 cards, where the training process takes about 40 hours.
Software Dependencies No Our model is implemented with Py Torch (42). While PyTorch is mentioned, a specific version number is not provided, nor are other software dependencies with their versions.
Experiment Setup Yes We adopt the Adam W optimizer with parameters β1 = 0.9, β1 = 0.99, and weight decay of 0.05. We set the batch size of 8 and trained the framework for 50 epochs on our large-scale synthetic dataset. The initial learning rate is set to 1 10 4 and is halved every 10 epochs. During the training stage, We randomly crop the input images with the resolution of 512 512 to patches of 128 128 for augmentation. The hyperparameters λm and λn are set to 0.1 and 1, respectively.