Pix2Poly

https://github.com/yeshwanth95/pix2poly
Admin
Pix2Poly is an end-to-end differentiable deep neural network that directly extracts polygonal...
Tokens:5,110
Snippets:42
Trust Score:7.3
License:MIT
Update:3 weeks ago
Show doc for...
Tokens:
Raw
### Train with torchrun

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Start training with distributed data parallelism. Replace `<num GPUs>` with the number of GPUs to use.

```shell
torchrun --nproc_per_node=<num GPUs> train_ddp.py
```

--------------------------------

### Run inria_to_coco.py to prepare INRIA train/val splits

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to convert INRIA raw tiles and ground truths into MS COCO format. The first 5 tiles of each city are kept as validation split. Requires the pix2poly_env environment.

```shell
# with pix2poly_env
python inria_to_coco.py
```

--------------------------------

### Run evaluate_mass_roads_predictions.py for Massachusetts Roads

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Computes evaluation metrics for raster predictions on the Massachusetts Roads Dataset. Requires ground truth directory and predicted raster masks folder.

```bash
python evaluate_mass_roads_predictions.py --gt-dir data/mass_roads_1500/test/map --dt-dir path/to/predicted/raster/masks/folder
```

--------------------------------

### Run spacenet_to_coco.py to prepare SpaceNet train/val splits

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to convert SpaceNet raw tiles and ground truths into MS COCO format. The first 15% tiles are kept as validation split. Requires the pix2poly_env environment.

```shell
# with pix2poly_env
python spacenet_to_coco.py
```

--------------------------------

### Run whu_buildings_to_coco.py to prepare WHU Buildings splits

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to convert WHU Buildings raw tiles and ground truths (512x512) into MS COCO format for train, validation, and test splits. Requires the pix2poly_env environment.

```shell
# with pix2poly_env
python whu_buildings_to_coco.py
```

--------------------------------

### Download Massachusetts Roads dataset

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to download the Massachusetts Roads Dataset using the provided shell script.

```shell
./download_mass_roads_dataset.sh
```

--------------------------------

### Run evaluation.py for building datasets

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Computes evaluation metrics for predicted building footprint files. The --eval-type argument accepts one of: ciou, angle, polis, or topdig.

```bash
python evaluation.py --gt-file path/to/ground/truth/annotation.json --dt-file path/to/prediction.json --eval-type <metric_type>
```

--------------------------------

### INRIA dataset directory structure

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Required folder layout for the raw INRIA Aerial Image Labeling Dataset before running inria_to_coco.py.

```text
data/inria_raw/
├── test/
│   └── images/
└── train/
    ├── gt/
    └── images/
```

--------------------------------

### Predict on INRIA dataset

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Generate predictions for the INRIA dataset. Run the prediction script with the required arguments, then run the postprocessing script after adjusting input/output paths in lines 6-10.

```shell
python predict_inria_coco_val_set.py -d inria_coco_224_negAug \
                                     -e <path_to_exp_folder> \
                                     -c <checkpoint_name> \
                                     -o <Name of output subdirectory>
python postprocess_coco_parts.py  # change input and output paths in L006 to L010.
```

--------------------------------

### Convert SpaceNet 16-bit to 8-bit

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to convert pansharpened RGB image tiles from 16-bit to 8-bit. Requires the gdal_env environment.

```shell
# with gdal_env
python spacenet_convert_16bit_to_8bit.py
```

--------------------------------

### Split Massachusetts Roads tiles into 224x224 patches

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to split the 1500x1500 tiles into 224x224 overlapping patches. Requires the gdal_env environment.

```shell
# using gdal_env
python mass_roads_tiles_to_patches.py
```

--------------------------------

### Predict on WHU Buildings dataset

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Generate predictions for the WHU Buildings dataset. Run the prediction script with the required arguments, then run the postprocessing script after adjusting input/output paths in lines 6-10.

```shell
python predict_whu_buildings_coco_test_set.py -d whu_buildings_224_coco \
                                              -e <path_to_exp_folder> \
                                              -c <checkpoint_name> \
                                              -o <Name of output subdirectory>
python postprocess_coco_parts.py  # change input and output paths in L006 to L010.
```

--------------------------------

### Conda environment specification for Pix2Poly

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Conda environment specification for Pix2Poly. Use with `conda env create -f environment.yml` (or equivalent) to set up the required dependencies, including PyTorch 2.1.2, CUDA 11.8, timm, transformers, and various image-processing libraries.

```yaml
channels:
  - defaults
dependencies:
  - torchvision=0.16.2
  - pytorch=2.1.2
  - pytorch-cuda=11.8
  - torchaudio=2.1.2
  - timm=0.9.12
  - transformers=4.32.1
  - pycocotools=2.0.6
  - torchmetrics=1.2.1
  - tensorboard=2.15.1
  - pip:
    - albumentations==1.3.1
    - imageio==2.33.1
    - matplotlib-inline==0.1.6
    - opencv-python-headless==4.8.1.78
    - scikit-image==0.22.0
    - scikit-learn==1.3.2
    - scipy==1.11.4
    - shapely==2.0.4
```

--------------------------------

### SpaceNet 2 Vegas dataset directory structure

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Required folder layout for the raw SpaceNet 2 Vegas subset before running the conversion scripts.

```text
data/AOI_2_Vegas_Train/
└── geojson/
    └── buildings/
└── RGB-PanSharpen/
    ├── gt/
    └── images/
```

--------------------------------

### Generate vector annotations for Massachusetts Roads patches

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Two commands to generate vector annotation files for the patches: first clip tile vectors, then convert world coordinates to pixel coordinates. Both require the gdal_env environment.

```shell
# using gdal_env
python mass_roads_clip_tile_vectors.py
python mass_roads_world_to_pixel_coords.py
```

--------------------------------

### WHU Buildings dataset directory structure

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Required folder layout for the raw WHU Buildings aerial imagery dataset (0.2 meter split) before running whu_buildings_to_coco.py.

```text
data/WHU_aerial_0.2/
├── test/
│   ├── image/
│   └── label/
├── train/
│   ├── image/
│   └── label/
└── val/
    ├── image/
    └── label/
```

--------------------------------

### Predict on Spacenet 2 dataset

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Generate predictions for the Spacenet 2 dataset. Run the prediction script with the required arguments, then run the postprocessing script after adjusting input/output paths in lines 6-10.

```shell
python predict_spacenet_coco_val_set.py -d spacenet_coco \
                                        -e <path_to_exp_folder> \
                                        -c <checkpoint_name> \
                                        -o <Name of output subdirectory>
python postprocess_coco_parts.py  # change input and output paths in L006 to L010.
```

--------------------------------

### Predict on Massachusetts Roads dataset

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

Generate predictions for the Massachusetts Roads dataset. Use `-s` to specify the split ('test' or 'val'), and adjust `--batch_size` according to available resources.

```shell
python predict_mass_roads_test_set.py -e <path_to_exp_folder> \
-c <checkpoint_name> \
-s <split> \  # 'test' or 'val'
--img_size 224 \
--input_size 224 \
--batch_size 24 \  # modify according to resources
```

--------------------------------

### Clip Massachusetts Roads shapefile to raster extents

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to generate vector annotations for each tile by clipping the roads geojson to the corresponding raster extents. Requires the gdal_env environment.

```shell
# using gdal_env
python mass_roads_clip_shapefile.py
```

--------------------------------

### Convert SpaceNet geojson to pixel coordinates

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Command to convert geojson annotations from world space coordinates to pixel space coordinates. Requires the gdal_env environment.

```shell
# with gdal_env
python spacenet_world_to_pixel_coords.py
```

--------------------------------

### Massachusetts Roads 224x224 directory structure

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Final directory structure of the processed 224x224 patches of the Massachusetts Roads Dataset, used for training Pix2Poly.

```text
data/mass_roads_224/
├── test/
│   ├── map/
│   ├── pixel_annotations/
│   ├── sat/
│   └── shape/
├── train/
│   ├── map/
│   ├── pixel_annotations/
│   ├── sat/
│   └── shape/
└── valid/
    ├── map/
    ├── pixel_annotations/
    ├── sat/
    └── shape/
```

--------------------------------

### Massachusetts Roads 1500x1500 directory structure

Source: https://github.com/yeshwanth95/pix2poly/blob/master/data_preprocess/README.md

Directory structure resulting from clipping the shapefile, containing 1500x1500 tiles of the Massachusetts Roads Dataset.

```text
data/mass_roads_1500/
├── test/
│   ├── map/
│   ├── sat/
│   └── shape/
├── train/
│   ├── map/
│   ├── sat/
│   └── shape/
└── valid/
    ├── map/
    ├── sat/
    └── shape/
```

--------------------------------

### Cite Pix2Poly paper (WACV 2025)

Source: https://github.com/yeshwanth95/pix2poly/blob/master/README.md

BibTeX citation for the Pix2Poly paper published at WACV 2025. Use this when citing the work.

```bibtex
@InProceedings{Adimoolam_2025_WACV,
    author    = {Adimoolam, Yeshwanth Kumar and Poullis, Charalambos and Averkiou, Melinos},
    title     = {Pix2Poly: A Sequence Prediction Method for End-to-End Polygonal Building Footprint Extraction from Remote Sensing Imagery},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
    month     = {February},
    year      = {2025},
    pages     = {8473-8482}
}
```
Tokens:
Raw