How kyotovision-public/fumet works
FUMET is a research implementation for training monocular road-scene depth estimators to recover metric scale from dashcam video. The principal workflow is to configure a KITTI or Cityscapes run, load temporal images and vehicle/road annotations, predict depth and pose, apply geometric and camera-height metric training, and save model outputs; separate scripts evaluate depth estimates. The graph preserves the supported encoder/decoder families and dataset-specific training branches. Edges use README statements or SOURCE INDEX references; unsampled implementation details are not inferred.
Main entry points: trainer_base.py Training Core · trainer_kitti.py KITTI Metric Trainer · trainer_cityscapes.py Cityscapes Metric Trainer · evaluate_depth.py Depth Evaluation
Components 26
Run Configuration
- Training Options
options.py - KITTI Trainer Entry
train_kitti.py - Cityscapes Trainer Entry
train_cityscapes.py
Dataset Pipeline
- KITTI Dataset
datasets/kitti.py - Cityscapes Dataset
datasets/cityscapes.py - Shared Data Loading
datasets/base.py - KITTI Geometry
kitti_utils.py
Metric Depth Training
- KITTI Metric Trainer
trainer_kitti.py - Cityscapes Metric Trainer
trainer_cityscapes.py - Training Core
trainer_base.py - Metric Geometry
utils.py - Depth Geometry Layers
layers.py
Model Architectures
- ResNet Encoder
networks/resnet_encoder.py - Depth Decoder Families
networks/depth_decoder.py - HR-Depth Decoder
networks/HR_depth_decoder.py - Lite-Mono Encoder
networks/litemono_encoder.py - Lite-Mono Decoder
networks/litemono_decoder.py - VAN Encoder
networks/van_encoder.py - VAN Backbone
networks/van.py - VAN Decoder
networks/van_decoder.py - Pose Networks
networks/pose_decoder.py
Depth Evaluation
- Depth Evaluation
evaluate_depth.py - Evaluation Datasets
datasets/base_eval.py - Evaluation Metrics
eval_utils.py
Other components
- Researcher
- Depth Model Weights
Connections 31
- Researcher → KITTI Trainer Entry: starts KITTI
README.md - Researcher → Cityscapes Trainer Entry: starts Cityscapes
README.md - KITTI Trainer Entry → Training Options: parses options
train_kitti.py - Cityscapes Trainer Entry → Training Options: parses options
train_cityscapes.py - KITTI Trainer Entry → KITTI Metric Trainer: starts training
train_kitti.py - Cityscapes Trainer Entry → Cityscapes Metric Trainer: starts training
train_cityscapes.py - KITTI Dataset → Shared Data Loading: inherits loading
datasets/kitti.py - Cityscapes Dataset → Shared Data Loading: inherits loading
datasets/cityscapes.py - KITTI Dataset → KITTI Geometry: generates depth
datasets/kitti.py - KITTI Metric Trainer → Training Core: extends core
trainer_kitti.py - Cityscapes Metric Trainer → Training Core: extends core
trainer_cityscapes.py - KITTI Metric Trainer → KITTI Dataset: loads training data
- Cityscapes Metric Trainer → Cityscapes Dataset: loads training data
- Training Core → Depth Geometry Layers: uses geometry
trainer_base.py - Training Core → Metric Geometry: uses metric utilities
trainer_base.py - KITTI Metric Trainer → Metric Geometry: computes metric loss
trainer_kitti.py - Cityscapes Metric Trainer → Metric Geometry: computes metric loss
trainer_cityscapes.py - Training Core → ResNet Encoder: constructs encoder
trainer_base.py - Training Core → Depth Decoder Families: constructs decoder
- Training Core → Pose Networks: constructs pose model
trainer_base.py - Training Core → HR-Depth Decoder: supports HR-Depth
README.md - Training Core → Lite-Mono Encoder: supports Lite-Mono
trainer_base.py - Training Core → Lite-Mono Decoder: supports Lite-Mono
- Training Core → VAN Encoder: supports VADepth
README.md - VAN Encoder → VAN Backbone: uses backbone
networks/van_encoder.py - Training Core → VAN Decoder: supports VADepth
- Training Core → Depth Model Weights: saves weights
README.md - Researcher → Depth Evaluation: evaluates model
evaluate_depth.py - Depth Evaluation → Evaluation Datasets: loads evaluation data
evaluate_depth.py - Depth Evaluation → Evaluation Metrics: calculates metrics
evaluate_depth.py - Depth Evaluation → Depth Model Weights: evaluates weights
README.md