#dynamic scenes

SC-Depth offers a robust, three-version framework for self-supervised monocular depth estimation. V1 features geometry consistency and self-discovered masks for reliable depth predictions. V2 enhances indoor scene accuracy with an auto-rectify network managing rotation challenges. V3 uses pretrained networks for improved depth accuracy in dynamic scenes. With support for multiple datasets and comprehensive scripts, it suits diverse research and application needs.

K-Planes offers an extensible radiance field model for a variety of static and dynamic datasets, applicable across different scene dimensions. This tool enables explicit modeling of radiance fields over space and time, improving the visualization of appearances in 3D spaces. The project's setup utilizes PyTorch within a conda environment, providing scripts for comprehensive training and evaluation. Integration with NerfAcc and NerfStudio libraries ensures an efficient training and visualization process. Delve into capabilities such as novel camera trajectory rendering and quality metric evaluation, beneficial for advancing neural rendering technologies.

The project introduces Neural-DynamicReconstruction (NDR), a method for reconstructing dynamic scenes with high fidelity using a monocular RGB-D camera. NDR is a template-free approach that accurately captures geometry, motion, and appearance. The project includes comprehensive setup guides, data conventions, and pre-processing steps. Built on NeuS technology, it integrates dependencies like PyMCubes and opencv_python for optimal results. It provides tools for training, evaluating models, and rendering geometries. Presented at NeurIPS 2022, this research advances dynamic scene visualization, serving both academic and industrial purposes.