cvlab-uos/viewsplat

How cvlab-uos/viewsplat works

ViewSplat is a feed-forward novel-view synthesis system that predicts and refines 3D Gaussian attributes for a requested camera pose. The principal workflow is configured in src/main.py: it constructs the model and data module, then trains or tests them; the model wrapper coordinates encoding, rendering, and losses. Separate capabilities include pose evaluation and in-the-wild point-cloud/video visualization. Edges are limited to relationships supported by the README, excerpts, or SOURCE INDEX; unsampled subsystem links are omitted where callers are not evidenced.

Main entry points: validate_in_the_wild.py Wild-Scene Validation · encoder_spf_viewsplat.py ViewSplat Encoder · main.py Train / Test · model_wrapper.py Model Wrapper

Components 21

Training Workflow

Train / Test src/main.py
Configuration src/config.py
Data Loading src/dataset/data_module.py
RE10K Dataset src/dataset/dataset_re10k.py
View Sampling src/dataset/view_sampler/view_sampler_bounded.py
Model Wrapper src/model/model_wrapper.py
Training Losses src/loss/loss_ssim.py

Scene Reconstruction

ViewSplat Encoder src/model/encoder/encoder_spf_viewsplat.py
Feature Backbones src/model/encoder/backbone
Gaussian Adapter src/model/encoder/common/gaussian_adapter.py
Camera Geometry src/geometry/projection.py
Gaussian Renderer src/model/decoder/decoder_splatting_cuda.py

Evaluation

Pose Evaluation Entry src/eval_pose.py
Pose Evaluator src/evaluation/pose_evaluator.py
Image Metrics src/evaluation/metrics.py
Evaluation Indices src/evaluation/evaluation_index_generator.py

Visualization

Wild-Scene Validation src/paper/validate_in_the_wild.py
Point-Cloud Export src/model/ply_export.py
Result Visualization src/visualization/layout.py
Image I/O src/misc/image_io.py

Other components

Researcher

Connections 24

Researcher → Train / Test: runs README.md
Train / Test → Configuration: loads configuration src/main.py
Train / Test → Data Loading: constructs data src/main.py
Train / Test → Model Wrapper: trains or tests src/main.py
Data Loading → RE10K Dataset: gets datasets src/dataset/data_module.py
Data Loading → View Sampling: configures sampling
Model Wrapper → ViewSplat Encoder: runs encoder src/model/model_wrapper.py
Model Wrapper → Gaussian Renderer: renders predictions
Model Wrapper → Training Losses: computes losses src/model/model_wrapper.py
ViewSplat Encoder → Feature Backbones: extracts features src/model/encoder/encoder_spf_viewsplat.py
ViewSplat Encoder → Gaussian Adapter: adapts Gaussians src/model/encoder/encoder_spf_viewsplat.py
ViewSplat Encoder → Camera Geometry: projects geometry src/model/encoder/encoder_spf_viewsplat.py
Gaussian Renderer → Camera Geometry: uses projection
Pose Evaluation Entry → Pose Evaluator: runs evaluation src/eval_pose.py
Pose Evaluation Entry → Data Loading: loads evaluation data src/eval_pose.py
Pose Evaluator → Image Metrics: uses metrics src/evaluation/pose_evaluator.py
Researcher → Pose Evaluation Entry: evaluates poses src/eval_pose.py
Evaluation Indices → Camera Geometry: computes geometry src/evaluation/evaluation_index_generator.py
Evaluation Indices → Result Visualization: lays out previews src/evaluation/evaluation_index_generator.py
Wild-Scene Validation → ViewSplat Encoder: builds reconstruction src/paper/validate_in_the_wild.py
Wild-Scene Validation → Gaussian Renderer: renders views src/paper/validate_in_the_wild.py
Wild-Scene Validation → Point-Cloud Export: exports point cloud src/paper/validate_in_the_wild.py
Wild-Scene Validation → Image I/O: writes images src/paper/validate_in_the_wild.py
Wild-Scene Validation → Result Visualization: creates visual outputs src/paper/validate_in_the_wild.py

Drawn by GitDiagram on November 13, 2026 from the repository's file tree, README and source files. Read as Markdown ·