How deepti0611/spacenet-off-nadir works
This repository is a research workflow for semantic segmentation of satellite imagery, particularly building masks. Its main stages are dataset construction and splitting, raster/mask loading and augmentation, model training, test evaluation, and result analysis. The tree also shows multiple model and training variants. Source excerpts establish a few training/evaluation dependencies; the README is empty, and many preparation and variant files were not sampled, so their connections are not asserted.
Main entry points: pytorch_dataset.py Torch dataset · train_deeplabv3plus_cbam_focal.py DeepLab training · train_unet_resnet34.py U-Net training · segformer SegFormer training
Components 20
Dataset preparation
- Dataset building
data_preparation/build_dataset.py - Dataset splitting
data_preparation/make_final_split.py - Dataset verification
checks/verify_dataset.py
Data loading
- Torch dataset
ml_pipeline/pytorch_dataset.py - Root dataset
pytorch_dataset.py - Dataset augmentation
ml_pipeline
Segmentation models
- DeepLabV3+
models/deeplabv3plus_resnet50/deeplabv3plus_resnet50_model.py - DeepLab CBAM
models/deeplabv3plus_cbam_bce_dice/deeplabv3plus_cbam_model.py - DeepLab training
models/deeplabv3plus_cbam_focal/train_deeplabv3plus_cbam_focal.py - U-Net model
models/unet_resnet34/unet_resnet34_model.py - U-Net training
models/unet_resnet34/train_unet_resnet34.py - SegFormer model
models/segformer/segformer_model.py - SegFormer training
models/segformer - SegFormer variants
models/segformer_cbam_bce_dice - DeepLab variants
models/deeplabv3plus_resnet50_partial
Evaluation and analysis
- SegFormer evaluation
models/segformer/evaluate_segformer.py - Variant evaluation
models/segformer_perspective/evaluate_segformer_perspective.py - Model analysis
analysis/segformer/generate_all_comparison_plots.py - Dataset analysis
analysis/analysis/building_area_analysis.py
Other components
- Researcher
Connections 14
- Researcher → Dataset building: prepares data
- DeepLab training → Torch dataset: loads samples
models/deeplabv3plus_cbam_focal/train_deeplabv3plus_cbam_focal.py - DeepLab training → DeepLab CBAM: creates model
models/deeplabv3plus_cbam_focal/train_deeplabv3plus_cbam_focal.py - U-Net training → Root dataset: loads samples
models/unet_resnet34/train_unet_resnet34.py - U-Net training → U-Net model: creates model
models/unet_resnet34/train_unet_resnet34.py - SegFormer training → Torch dataset: uses dataset
- SegFormer training → SegFormer model: uses model
- DeepLab variants → Torch dataset: uses dataset
models/deeplabv3plus_resnet50_partial/deeplabv3plus_resnet50_partial_train.py - DeepLab variants → DeepLabV3+: uses model
models/deeplabv3plus_resnet50_partial/deeplabv3plus_resnet50_partial_train.py - SegFormer variants → Torch dataset: uses dataset
- SegFormer variants → SegFormer model: uses model
- SegFormer evaluation → Root dataset: loads test data
models/segformer/evaluate_segformer.py - Variant evaluation → Torch dataset: uses dataset
models/segformer_perspective/evaluate_segformer_perspective.py - Researcher → Model analysis: compares results