How sdmuhsin/dlo-segmentation works
This repository is a deformable-linear-object (DLO) semantic-segmentation research pipeline: it renders labelled RGB-D imagery, prepares three-class training data, trains segmentation models, and evaluates their predictions. Its documented reproduction flow runs from assets through rendering, staging and label re-encoding to training and evaluation. Separate branches cover real-photo dataset adaptation and video-data generation. The graph uses README statements and the SOURCE INDEX for relationships; many implementation details—especially rendering and inference variants—were not sampled, so those links are omitted or marked unverified.
Main entry points: builder.py Model builder · train_dformer_v2_dlo.py DFormer training · reencode_labels_3way.py Three-class labels · build_p44_connpaste_dataset.py Composite dataset assembly
Components 23
Data preparation
- Synthetic renderer
src/render_full_dataset.py - Video dataset conversion
src/convert_to_video_dataset.py - Point-cloud RGB-D
src/pcl_to_rgbd.py - Texture mapping
src/texture_mapping.py - Dataset staging
src/prepare_dformer_data.py - Three-class labels
src/reencode_labels_3way.py
Models and training
- RGB-D training
src/train_rgbd_seg.py - RGB training
src/train_rgb_only_sota.py - DFormer training
src/train_dformer_v2_dlo.py - Training runtime
src/dformer/utils/train.py - Model builder
src/dformer/models/builder.py - DFormer encoder
src/dformer/models/encoders/DFormer.py - DFormer v2 encoder
src/dformer/models/encoders/DFormerv2.py - Segmentation decoders
src/dformer/models/decoders/nl_head.py
Evaluation and inference
- Three-way evaluation
src/eval_3way_final.py - Segmentation inference
src/dformer/mmseg/apis/inference.py
Real-world adaptation
- Electric Wires dataset
dataset_generation/realworld_dataset.py - Photo composites
src/composite_connector_positives.py - Composite dataset assembly
src/build_p44_connpaste_dataset.py - Composite validation
src/validate_p44_composites.py
Other components
- Researcher
- Input assets
- Camera footage
Connections 24
- Researcher → Input assets: fetches
README.md - Input assets → Synthetic renderer: supplies inputs
README.md - Researcher → Synthetic renderer: runs
README.md - Synthetic renderer → Dataset staging: renders RGB-D labels
README.md - Dataset staging → Three-class labels: stages for training
README.md - Three-class labels → RGB-D training: provides dataset
README.md - Researcher → RGB-D training: trains model
README.md - Researcher → Three-way evaluation: evaluates checkpoint
README.md - Three-class labels → Three-way evaluation: provides evaluation data
README.md - DFormer training → Model builder: imports
src/train_dformer_v2_dlo.py - DFormer training → RGB-D training: imports
src/train_dformer_v2_dlo.py - DFormer training → RGB training: imports
src/train_dformer_v2_dlo.py - RGB training → Electric Wires dataset: imports
src/train_rgb_only_sota.py - Video dataset conversion → Point-cloud RGB-D: imports
src/convert_to_video_dataset.py - Video dataset conversion → Texture mapping: imports
src/convert_to_video_dataset.py - Photo composites → Composite dataset assembly: builds paste set
README.md - Composite dataset assembly → Composite validation: validated by
README.md - Composite dataset assembly → RGB training: trains on
- Camera footage → Three-way evaluation: supplies footage
- Segmentation inference → Camera footage: predicts on
- Training runtime → Model builder: builds model
src/dformer/utils/train.py - Model builder → DFormer encoder: selects encoder
src/dformer/models/builder.py - Model builder → DFormer v2 encoder: selects encoder
src/dformer/models/builder.py - Model builder → Segmentation decoders: selects decoder
src/dformer/models/builder.py