How scnu-rislab/mf-mos works
MF-MOS is a SemanticKITTI moving-object segmentation system. Its documented workflow derives temporal residual maps from LiDAR sequences, loads scans and labels for training or inference, predicts moving-object classes, and evaluates or visualizes those predictions. Training has a base-model stage and an optional SIEM point-refinement stage. The excerpts confirm dataset parsing, training orchestration, refinement construction, residual generation, and evaluation; inference and some model wiring are represented only at the subsystem level where the tree or README supports them.
Main entry points: trainer.py Training loop · parser.py KITTI parser · auto_gen_residual_images_mp.py Residual generation · train.py Base training entry
Components 16
Data preparation
- Residual generation
utils/auto_gen_residual_images_mp.py - KITTI parser
common/dataset/kitti/parser.py
Model and training
- Base training entry
train.py - Training loop
modules/trainer.py - MF-MOS model
modules/MFMOS.py - Refinement training
train_2stage.py - Refinement trainer
modules/trainer_refine.py - SIEM point network
modules/PointRefine/spvcnn.py
Prediction and analysis
- Inference
infer.py - MOS evaluation
utils/evaluate_mos.py - Result visualization
utils/visualize_mos.py - Sequence video
utils/viz_seqVideo.py
Scan and metrics
- Scan projection
common/laserscan.py - Segmentation losses
modules/loss
Other components
- Researcher
- SemanticKITTI data
Connections 18
- Researcher → Residual generation: generates maps
- SemanticKITTI data → Residual generation: reads sequences
- Residual generation → KITTI parser: provides residual maps
- SemanticKITTI data → KITTI parser: provides scans and labels
- KITTI parser → Scan projection: projects scans
- Researcher → Base training entry: starts training
- Base training entry → Training loop: creates trainer
- Training loop → KITTI parser: loads dataset
- Training loop → MF-MOS model: trains model
- Training loop → Segmentation losses: computes loss
- Researcher → Refinement training: starts refinement
- Refinement training → Refinement trainer: runs refinement
- Refinement trainer → SIEM point network: constructs network
- Researcher → Inference: requests predictions
- Researcher → MOS evaluation: evaluates predictions
- SemanticKITTI data → MOS evaluation: provides ground truth
- Researcher → Result visualization: visualizes results
- Researcher → Sequence video: views sequences