RayDF: Neural Ray-surface Distance Fields with Multi-view Consistency

Authors: Zhuoman Liu, Bo Yang, Yan Luximon, Ajay Kumar, Jinxi Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We extensively evaluate our method on three public datasets, demonstrating remarkable performance in 3D surface point reconstruction on both synthetic and challenging real-world 3D scenes, clearly surpassing existing coordinate-based and ray-based baselines. Most notably, our method achieves a 1000 faster speed than coordinate-based methods to render an 800 800 depth image, showing the superiority of our method for 3D shape representation.
Researcher Affiliation Academia Zhuoman Liu Bo Yang Yan Luximon Ajay Kumar Jinxi Li v LAR Group, The Hong Kong Polytechnic University {zhuo-man.liu, jinxi.li}@connect.polyu.hk bo.yang@polyu.edu.hk
Pseudocode No The paper does not contain a clearly labeled pseudocode or algorithm block.
Open Source Code Yes Our code and data are available at https://github.com/v LAR-group/Ray DF
Open Datasets Yes Our method is evaluated on three types of public datasets: 1) the object-level synthetic Blender dataset from the original Ne RF paper [47], 2) the scene-level synthetic DM-SR dataset from the recent DM-Ne RF paper [73], and 3) the scene-level real-world Scan Net dataset [16].
Dataset Splits No The paper specifies training and testing splits (e.g., '100 views for training and 200 novel views for testing' for Blender dataset) but does not explicitly mention a separate validation split.
Hardware Specification Yes we conduct a simple experiment to generate an 800 800 depth image on a computer with a single NVIDIA RTX 3090 GPU card and a CPU of AMD Ryzen 7.
Software Dependencies No The paper mentions using Adam optimizer and SIREN network architecture but does not specify version numbers for any software dependencies or libraries used for implementation.
Experiment Setup Yes We use a 13-layer SIREN [63] with 1024 hidden units to learn a ray-surface distance field. We set the batch size as 8192 and train the network for 10 epochs with the Adam [35] optimizer and a cosine annealing strategy [44] with the learning rate initialized as 10 5 and decayed to 10 8. For the auxiliary network, we use a 7-layer SIREN with 512 hidden units, with batch sizes of 2048 for Blender and ScanNet, and 1024 for DM-SR, training for 5 epochs with Adam and a cycle annealing strategy with max learning rate 10 4.