giswqs/nwpu-vhr10-maskrcnn

Mask R-CNN for Multi-Class Object Detection on NWPU-VHR-10

This model is a Mask R-CNN with a ResNet-50 FPN backbone, trained for multi-class object detection on very high resolution (VHR) remote sensing imagery using the NWPU-VHR-10 benchmark dataset.

Model Description

• Architecture: Mask R-CNN with ResNet-50 FPN backbone (torchvision maskrcnn_resnet50_fpn)
• Task: Multi-class object detection and instance segmentation
• Domain: Very high resolution (VHR) remote sensing imagery
• Training Framework: PyTorch + torchvision
• Package: Trained using geoai-py

Training Data

The NWPU-VHR-10 dataset was constructed by researchers at Northwestern Polytechnical University (NWPU). It contains:

• 800 VHR remote sensing images (650 positive with annotations, 150 negative)
• 3,775 annotated object instances across 10 classes
• Images cropped from Google Earth and the Vaihingen dataset
• Manually annotated by domain experts

Classes

ID	Class	ID	Class
1	airplane	6	basketball_court
2	ship	7	ground_track_field
3	storage_tank	8	harbor
4	baseball_diamond	9	bridge
5	tennis_court	10	vehicle

Data Split

Split	Images	Purpose
Train	509 (85%)	Model training
Val	128 (15%)	Evaluation

Training Details

Parameter	Value
Backbone	ResNet-50 FPN (ImageNet pretrained)
Optimizer	SGD (momentum=0.9, weight_decay=0.0005)
Learning Rate	0.005 with StepLR (step=3, gamma=0.1)
Batch Size	4
Epochs	20
Input Channels	3 (RGB)
Seed	42

Evaluation Results

Evaluated on the validation set (128 images) using standard COCO detection metrics:

Metric	Value
mAP@0.5	0.709
mAP@0.75	0.518
mAP@[0.5:0.95]	0.459

Per-Class AP@0.5

Class	AP@0.5	Class	AP@0.5
tennis_court	0.902	harbor	0.695
basketball_court	0.878	bridge	0.678
baseball_diamond	0.821	storage_tank	0.591
ground_track_field	0.807	vehicle	0.532
airplane	0.713	ship	0.470

Usage

With geoai

import geoai

# Run inference on a remote sensing image
result_path, inference_time, detections = geoai.multiclass_detection(
    input_path="your_image.tif",
    output_path="detections.tif",
    model_path="best_model.pth",
    num_classes=11,
    class_names=[
        "background", "airplane", "ship", "storage_tank",
        "baseball_diamond", "tennis_court", "basketball_court",
        "ground_track_field", "harbor", "bridge", "vehicle",
    ],
    window_size=512,
    overlap=256,
    confidence_threshold=0.5,
)

# Visualize results
geoai.visualize_multiclass_detections(
    image_path="your_image.tif",
    detections=detections,
    class_names=[
        "background", "airplane", "ship", "storage_tank",
        "baseball_diamond", "tennis_court", "basketball_court",
        "ground_track_field", "harbor", "bridge", "vehicle",
    ],
)

With PyTorch directly

import torch
import torchvision
from torchvision.models.detection import maskrcnn_resnet50_fpn

# Load model
model = maskrcnn_resnet50_fpn(weights=None, num_classes=11)
checkpoint = torch.load("best_model.pth", map_location="cpu")
model.load_state_dict(checkpoint)
model.eval()

# Run inference
from PIL import Image
from torchvision import transforms

img = Image.open("your_image.jpg").convert("RGB")
img_tensor = transforms.ToTensor()(img).unsqueeze(0)

with torch.no_grad():
    predictions = model(img_tensor)

# predictions[0] contains 'boxes', 'labels', 'scores', 'masks'

Files

File	Description
best_model.pth	Best model weights (by validation IoU)
class_info.json	Class names and number of classes
training_summary.txt	Training configuration summary

Notebook Example

For a complete end-to-end example including dataset download, training, evaluation, and inference, see the NWPU-VHR-10 Object Detection notebook.

Citation

If you use this model, please cite the NWPU-VHR-10 dataset:

@article{cheng2014multi,
  title={Multi-class geospatial object detection and geographic image classification based on collection of part detectors},
  author={Cheng, Gong and Han, Junwei and Zhou, Peicheng and Guo, Lei},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={98},
  pages={119--132},
  year={2014}
}

@article{cheng2016survey,
  title={A survey on object detection in optical remote sensing images},
  author={Cheng, Gong and Han, Junwei},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={117},
  pages={11--28},
  year={2016}
}

License

This model is released under the MIT License. The NWPU-VHR-10 dataset is for research purposes only.