Object Detection (OD) in aerial images has gained much attention due to its applications in search and rescue, town planning, and agriculture yield prediction etc. Recently introduced large-scale aerial images dataset, iSAID has enabled the researchers to advance the OD tasks on satellite images. Unfortunately, the available OD pipelines and ready-to-train architectures are well-tailored and configured to be used with tasks dealing with natural images. In this work, we study that directly using the available object detectors, specifically the vanilla Faster RCNN with FPN is sub-optimal for aerial OD. To help improve its performance, we tailor the Faster R-CNN architecture and propose several modifications including changes in architecture in different blocks of detector, training & transfer learning strategies, loss formulations, and other pre-post processing techniques. By adopting the proposed modifications on top of the vanilla Faster-RCNN, we push the performance of the model and achieve an absolute gain of 4.44 AP over the vanilla Faster R-CNN on the iSAID validation set.

This repository contains the code files for reproducing the main experiments mentioned in our project. Moreover, this repo supports the use of third party backbones to be integrated with the Faster R-CNN object detectors including SWIN, ConvNext and timm backbones.

Complete technical report can be viewed here.

We have tested this code on Ubuntu 20.04 LTS with Python 3.8. This repo is heavily built on Detectron2. Follow the instructions below to setup the environment and install the dependencies.

conda create -n detectron_OD python=3.8
conda activate detectron_OD
# Install torch and torchvision
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
# Install Detectron2 (for more details visit : https://detectron2.readthedocs.io/en/latest/tutorials/install.html)
python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
# Install matplotlib for visualizations
python -m pip install -U matplotlib

Now clone this repository:

git clone https://github.com/MUKhattak/OD-Satellite-iSAID.git
cd OD-Satellite-iSAID/

We use iSAID dataset[1]. You can download the dataset from here. Our code expects the dataset folder to have the following structure,

isaid_dataset_root-folder/
└─ train
    ├─ images
        ├─ P1207_1800_2600_1200_2000.png
        ├─ P1207_1800_2600_1800_2600.png
        ├─ .......
    ├─ instancesonly_filtered_train.json
└─ val
    ├─ images
        ├─ P1557_3000_3800_0_800.png
        ├─ .......
    ├─ instancesonly_filtered_val.json

We implement our code as seperate project in detectron2, so first cd to the project folder by running the following command.

$ cd projects/OD_satellite_iSAID/

To train a vanilla Faster R-CNN with FPN-R101 backbone, run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/faster_rcnn_R_101_FPN_3x.yaml --isaid-path /path/to/isaid/root/folder

To train a Faster R-CNN with ConvNext-Base backbone, run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/faster_rcnn_convnext_base_FPN_3x.yaml --isaid-path /path/to/isaid/root/folder

To train a Faster R-CNN with SWIN-Base backbone, run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/faster_rcnn_swin_base_3x_FPN.yaml --isaid-path /path/to/isaid/root/folder

To train a Faster R-CNN with timm-based ResNet-BiT backbone, run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/fastrcnn_timm_backbone.yaml --isaid-path /path/to/isaid/root/folder

Note: To use any other backbone from timm (pytorch image models), you may need to change slightly change the config file and timm_backbone.pyhere depending on the number of feature maps that can be obtained from the respective timm backbone.

To train a Faster R-CNN with Deformable ResNet-101 backbone,run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/faster_rcnn_deformable_resnet101.yaml --isaid-path /path/to/isaid/root/folder

To train a Faster R-CNN FPN-R101 with federated loss (+ sigmoid cross-entropy) , run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/faster_rcnn_fed_loss.yaml --isaid-path /path/to/isaid/root/folder

To train a Faster R-CNN FPN-R101 with focal loss, run the following command

$ python plain_train_net.py --config-file ./configs/derived_configs/faster_rcnn_focal_loss.yaml --isaid-path /path/to/isaid/root/folder

We have also provided additional config files which can be explored at projects/OD_satellite_iSAID/configs/derived_configs Further, other hyperparameters such as learning rate, iteration number can be modified in the config file.

You can use the different configurations together (e.g using specific backbone and loss function together) by merging the respective config files, and use the new config file for training.

We also explore other architectural and training modifications in the vanilla Faster R-CNN detector, please refer to our report and base config file for additional details.

To evaluate a trained model on the iSAID validation set, run the following command

$ python plain_train_net.py --config-file ./path/to/custom/config/file --eval-only --isaid-path /path/to/isaid/root/folder

We also provide a jupyter notebook for visualizing the GT and predicted bounding boxes on respective images for qualitative results. Please refer to this path to view the notebook projects/OD_satellite_iSAID/Detectron2_Faster-RCNN_iSAID.ipynb (or click here).

[1]Waqas Zamir, Syed, et al. "isaid: A large-scale dataset for instance segmentation in aerial images." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2019.