How wzzheng/dvgt works
DVGT is an autonomous-driving vision and geometry framework: users provide multi-view driving images, and DVGT-1 predicts metric-scale dense point maps while DVGT-2 adds online reconstruction and trajectory planning. The repository also supports dataset preparation, model training, metric-based evaluation, and visualization. Edges reflect README statements or shown source/index references; some training and dataset wiring is configurable or unsampled, so those links are omitted rather than inferred.
Main entry points: dvgt1.py DVGT-1 · dvgt2.py DVGT-2 · scene_dataset.py Scene dataset · metric_utils.py Metric dispatch
Components 25
Data preparation
- Annotation generation
tools/data_process/main.py - Dataset loaders
tools/data_process/dataloaders - Depth alignment
tools/data_process/utils/alignment.py - Scene dataset
dvgt/datasets/scene_dataset.py - Sequence transforms
dvgt/datasets/transforms
Geometry models
- DVGT-1
dvgt/models/architectures/dvgt1.py - DVGT-2
dvgt/models/architectures/dvgt2.py - DVGT-1 aggregator
dvgt/models/backbones/dvgt1_aggregator.py - DVGT-2 aggregator
dvgt/models/backbones/dvgt2_aggregator.py - Dense point head
dvgt/models/heads/dpt_head.py - Pose heads
dvgt/models/heads/dvgt1_ego_pose_head.py - Trajectory decoder
dvgt/models/heads/trajectory_pose_decoder.py - Attention blocks
dvgt/models/layers/block.py
Training and evaluation
- Training engine
dvgt/engine/trainer.py - Evaluation engine
dvgt/engine/evaluator.py
Metrics and geometry
- Metric dispatch
dvgt/evaluation/utils/metric_utils.py - Depth metrics
dvgt/evaluation/metrics/depth_metrics.py - Point cloud metrics
dvgt/evaluation/metrics/point_cloud_metrics.py - Evaluation geometry
dvgt/evaluation/utils/geometry.py - Pose geometry
dvgt/utils/geometry.py
Visualization
- Point cloud viewer
dvgt/visualization/visual.py - Visualization utilities
dvgt/visualization/utils.py
Other components
- Driving data user
- Multi-view images
- Reconstruction and plans
Connections 24
- Driving data user → Multi-view images: provides
README.md - Multi-view images → DVGT-1: infer from
README.md - Multi-view images → DVGT-2: infer from
README.md - DVGT-1 → Reconstruction and plans: predicts points
README.md - DVGT-2 → Reconstruction and plans: reconstructs and plans
README.md - Annotation generation → Dataset loaders: gets dataset
tools/data_process/main.py - Annotation generation → Depth alignment: references
tools/data_process/main.py - Scene dataset → Sequence transforms: uses
dvgt/datasets/scene_dataset.py - Scene dataset → Pose geometry: uses
dvgt/datasets/scene_dataset.py - Scene dataset → Multi-view images: reads
dvgt/datasets/scene_dataset.py - DVGT-1 → DVGT-1 aggregator: calls
dvgt/models/architectures/dvgt1.py - DVGT-1 → Dense point head: calls
dvgt/models/architectures/dvgt1.py - DVGT-1 → Pose heads: calls
dvgt/models/architectures/dvgt1.py - DVGT-2 → DVGT-2 aggregator: calls
dvgt/models/architectures/dvgt2.py - DVGT-2 → Dense point head: calls
dvgt/models/architectures/dvgt2.py - DVGT-2 aggregator → Attention blocks: uses blocks
dvgt/models/backbones/dvgt2_aggregator.py - Attention blocks → DVGT-2: supports
- Training engine → Pose geometry: imports
dvgt/engine/trainer.py - Evaluation engine → Metric dispatch: uses
dvgt/engine/evaluator.py - Metric dispatch → Depth metrics: dispatches
dvgt/evaluation/utils/metric_utils.py - Metric dispatch → Point cloud metrics: dispatches
dvgt/evaluation/utils/metric_utils.py - Evaluation geometry → Pose geometry: uses
dvgt/evaluation/utils/geometry.py - Point cloud viewer → Visualization utilities: uses
dvgt/visualization/visual.py - Point cloud viewer → Reconstruction and plans: displays
dvgt/visualization/visual.py