isbnet-building-instance-segmentation

KO | EN

gitlite — search

#python #llm #golang #go #android #linux #docker #macos #java #windows #kubernetes #vue #ios #openai

isbnet-building-instance-segmentation

(★ 12)

Building instance segmentation for aerial urban point clouds based on ISBNet, from a Computer Vision and Pattern Recognition final project.

isbnet-building-instance-segmentation Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ISBNet Building Instance Segmentation for Aerial Urban Point Clouds

This repository contains a lightweight, reproducible project for **building-only instance segmentation** on aerial urban point clouds. It was organized from a Sun Yat-sen University **Computer Vision and Pattern Recognition** final project and is built on top of [ISBNet](https://github.com/VinAIResearch/ISBNet).

The project focuses on converting custom building annotations into ISBNet format, fine-tuning an ISBNet building instance model, and improving instance-level results with overlapping blocks, cached candidates, and cross-block instance merging.

> Raw point clouds, manual annotation data, personal course records, large checkpoints, and full CloudCompare point-cloud visualizations are not included.

## Highlights

- Converts manually annotated building txt files into ISBNet-compatible `.pth` scenes.
- Treats each building txt as one instance and the largest-numbered txt as background.
- Supports building-only instance segmentation with `semantic_label=1` for buildings and `instance_label=-100` for background.
- Uses 50 m x 50 m overlapping XY blocks with 35 m stride.
- Caches block-level instance candidates and performs validation-selected cross-block graph merging.
- Exports point-level and instance-level metrics.
- Supports CloudCompare visualization export in `x y z r g b` format.

## Final Reported Result

Independent test group: `11-NW-10B`

| Metric | Value |
|---|---:|
| Instance Precision@0.5 | 0.6207 |
| Instance Recall@0.5 | 0.6102 |
| Instance F1@0.5 | 0.6154 |
| Matched instances@0.5 | 72 / 118 |
| Predicted instances | 116 |
| Matched instance mean IoU@0.5 | 0.7933 |
| Instance F1@0.25 | 0.7265 |

See `metrics/` for full per-scene and summary metrics.

## Repository Structure

```text
.
├── configs/              # Runtime configs and merge-search profiles
├── scripts/              # Data conversion, block preparation, evaluation, merge refinement
├── run/                  # Example end-to-end shell scripts
├── metrics/              # Final test metrics and validation search summary
├── examples/             # Selected visualization scene names
├── assets/               # Optional PNG/JPG screenshots for README/PPT
├── docs/                 # Method notes
├── requirements.txt
├── environment.yml
├── .gitignore
└── LICENSE
```

## Data Format

Raw manually annotated data should be organized by large groups and small scene folders. Each valid scene folder contains multiple txt files. Each txt row has:

```text
x y z r g b
```

Annotation convention:

- Each non-background txt file represents one building instance.
- The txt file with the largest numeric filename is background/non-building.
- Building points: `semantic_label=1`, continuous `instance_label=0,1,2,...`.
- Background points: `semantic_label=0`, `instance_label=-100`.

Converted scene format:

```python
torch.save((coords, colors, sem_labels, instance_labels), out_path)
```

## Usage

### 1. Prepare Custom Building Data

```bash
python scripts/prepare_our_building_finetune_data.py --overwrite
```

Adjust paths inside the script or pass arguments according to your local setup.

### 2. Train Overlapping-Block Model

```bash
CUDA_VISIBLE_DEVICES=0 bash run/run_instance_structure_all.sh
```

### 3. Cached Cross-Block Merge Refinement

```bash
CUDA_VISIBLE_DEVICES=0 bash run/run_merge_refinement_all.sh
```

This step caches block predictions, searches merge parameters on validation scenes, and evaluates once on the independent test group.

## Visualization Assets

For GitHub, this repository only keeps lightweight screenshots in `assets/`. The recommended five figures are:

1. Best qualitative result #1.
2. Best qualitative result #2.
3. Best qualitative result #3.
4. Original raw point-cloud scene.
5. Manual building-instance annotation, where only complete buildings are labeled.

The final CloudCompare txt files are intentionally not tracked in git because they are large. The exporter writes CloudCompare-friendly files:

```text
x y z r g b
```

Color convention:

- Gray: true background, predicted background (TN)
- Red: background predicted as building (FP)
- Blue: building missed by the model (FN)
- Other colors: correctly predicted building points (TP), with different colors for different predicted instances

For GitHub, it is recommended to include 2-4 small screenshots in `assets/`, not the full point-cloud txt files.

## Checkpoints

Large `.pth` checkpoints are not included in this repository. Put your checkpoint under your local ISBNet work directory and update the config paths accordingly.

## License

This project is released under the MIT License. Please also follow the license of the upstream ISBNet repository.

isbnet-building-instance-segmentation

isbnet-building-instance-segmentation Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

isbnet-building-instance-segmentation

isbnet-building-instance-segmentation Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README