How authoritywang/videogs works
Project Type and Purpose
• This is a research-oriented processing pipeline (not a web app) for generating, compressing, and viewing volumetric videos on mobile devices using 2D dynamic Gaussians.
• It comprises four main phases: Data Preprocessing, Model Training, Compression, and Rendering/Serving.
Break Out the Main Components
A. Data Sources & Storage
• Raw Datasets (ReRF, HiFi4G, HumanRF)
• File-based storage for “processed” datasets (datasets/…)
• Output folder (output/…)
• Configuration store (config/config_hash.json)
B. External Tools & Libraries
• COLMAP (for calibration & undistortion)
• NeuS2_K (key-frame point cloud generator)
• Submodules: diff-gaussian-rasterization, simple-knn
• LPIPS (perceptual loss)
C. Preprocessing Subsystem
• Scripts in preprocess/ (calibration.py, undistortion.py, colmap2k.py, hifi4g_process.py)
• Utilities from utils/ (camera_utils, general_utils)
• CLI entry point: none (invoked via explicit python commands)
D. Training Subsystem
• Entry scripts: train_sequence.py, train_dynamic.py, train_prune.py, train.py
• Core modules in scene/ (gaussian_model, global_t_field, dataset_readers)
• Loss modules: lpipsPyTorch + utils/loss_utils.py
• Arguments parsing: arguments/
• Output: Gaussian checkpoint folders + NeuS2 outputs
E. Compression & Serving Subsystem
• Scripts in compress/ (compress_ckpt_2_image_precompute.py, compress_image_2_video.py, decompress_video_2_ckpt.py)
• Viewer/rendering: gaussian_renderer/network_gui.py + run.py, convert.py
• Web Server (nginx) hosting the compressed videos + viewer front-end assets (assets/)
Architectural Patterns & Key Technologies
• Pipelines: clearly separate each phase into its own lane.
• Modular CLI: show Python scripts as rectangles calling into underlying modules.