RoIPoly

Environment Setup

1. Create Conda Environment (Recommended)

To begin, create and activate a new conda environment for this project:

conda create -n roipoly python=3.8
conda activate roipoly

2. Install PyTorch

Install the specific version of PyTorch compatible with CUDA 11.1:

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html

3. Install Detectron2

Follow the official Detectron2 installation guide for installation instructions.

If you encounter any issues during installation, you can directly download the detectron2 folder and place it under the RoIPoly directory. The final structure should look like this:

RoIPoly/
│
├── detectron2/
├── other files...

4. Install Boundary IoU API

Follow the official Boundary IoU API installation guide for installation instructions.

5. Install Required Packages

To install the remaining dependencies, run:

pip install -r requirements.txt

6. Compile Deformable-Attention Modules (from Deformable-DETR)

To compile the deformable attention modules, navigate to the ops directory and execute the following:

cd roipoly/ops
sh make.sh

Training the model

To train the model, use the following command:

CUDA_VISIBLE_DEVICES=<gpu_ids> python3 train_net.py --num-gpus <number_of_gpus> --config-file <path_to_config_file> --dataset-name <dataset_name> --train-json <path_to_annotation_json> --train-path <path_to_training_images>

Arguments:

• CUDA_VISIBLE_DEVICES (optional): This environment variable is used to specify which GPU(s) to use. For example, CUDA_VISIBLE_DEVICES=1 makes the second GPU (index 1) available for training.
  • If you want to use multiple GPUs, you can specify them like CUDA_VISIBLE_DEVICES=0,1,2.
  • If not set, the first available GPU will be used by default.
• --num-gpus: This argument specifies how many GPUs to use during training. Set --num-gpus <number_of_gpus> to define the number of GPUs. For example, --num-gpus 1 for single-GPU training and --num-gpus 2 for two GPUs.
• --config-file: Path to the configuration file that contains model-specific parameters, dataset paths, and other training settings. Replace <path_to_config_file> with the actual path to your configuration file.
• --dataset-name: The name of the dataset being used. Default is aicrowd_train, but you can change it to your own dataset name.
• --train-json: Path to the training dataset's annotation file in COCO format. Replace <path_to_annotation_json> with the path to your dataset's annotation JSON file.
• --train-path: Path to the folder containing training images. Replace <path_to_training_images> with the actual path to your dataset's images.

Example:

If you want to train a model using the Swin-Base Transformer as the backbone on the CrowdAI Small-Medium dataset, you can execute the following command:

CUDA_VISIBLE_DEVICES=1 python3 scripts/train_net.py --num-gpus 1 --config-file configs/roipoly.res50.34pro.aicrowd.yaml --dataset-name crowdai_train --train-json ./data/crowdai/train/annotation_sm_clean_us_index.json --train-path ./data/crowdai/train/images

🔍 Tips for Using This Repository

1. ROI Generation Strategy

   • During training, we use ground-truth bounding boxes to provide region proposals (RoIs).
   • During inference, RoIs are obtained from a pretrained object detector.

2. Polygon Size Partitioning (for CrowdAI Dataset)

   • In training, we split the CrowdAI dataset into small-medium polygons and large polygons.
   • This design is purely due to limited computational resources, not model performance considerations.
   • Without this split, we would need to set a very large num_proposal_vertices_per_polygon to handle complex large polygons, which would significantly increase memory and computation cost.
   • ⚠️ If you have sufficient computing resources, we do not recommend applying this split.

Inference

• scripts/evaluate.py

🚀 New Paper Alert

This repository is part of a broader research effort on vectorized object outline extraction for large-scale topographic map generation. Check out our latest publications:

• LDPoly: Latent diffusion for polygonal road outline extraction in large-scale topographic mapping
  ISPRS Journal of Photogrammetry and Remote Sensing, 2025
  📄 Paper (DOI) ｜ 💻 Code & Pretrained Models & Dataset
  → The first generative framework and benchmark for polygonal road outline extraction.