Remote-Sensing-Target-Localization

Highlights

We express the localization problem as a decomposable point querying process, concurrently addressing the counting issue. In this process, sparse input points can bifurcate into four new points when necessary. Our project is designed for aerial scenes in remote sensing, with the current focus on locating objects such as ships, crowds, and aircraft.

Installation

• Our project has no extra compiled components, minimal package dependencies, making the code straightforward and easy to use. Instructions for installing dependencies via conda are provided. First, clone the repository locally:

git clone https://github.com/tiny-smart/Remote-Sensing-Target-Counting-and-Localization.git

• Required packages:

torch
torchvision
numpy
opencv-python
scipy
matplotlib

• Then, install packages as:

pip install -r requirements.txt

Data Preparation

• Download crowd-counting datasets, e.g., ShanghaiTech.

• We expect the directory structure to be as follows:

PET
├── data
│    ├── ShanghaiTech
├── datasets
├── models
├── ...

• Download CARPK datasets, e.g., CARPK.

• We expect the directory structure to be as follows:

PET
├── data
│    ├── CARPK
├── datasets
├── models
├── ...

• I have transformed the CARPK dataset into a universal VOC dataset format. For unannotated images, you can use LabelImg (ps:c8lc) to standardize the annotations for the images. For annotated images, you can refer to 'Remote-Sensing-Target-Localization/util/convert_bbox_to_points.py' to convert existing data into the VOC dataset format.

• We expect the directory structure to be as follows:

• Alternatively, you can define the path of the dataset in datasets/init.py

• We also provide other remote sensing datasets (crowds, ships and vehicles) in consistent formats:

Dataset	Link
crowds	https://pan.baidu.com/s/1XVzgoyriQ4ibnc_cQM9v8A
vehicles	https://pan.baidu.com/s/1j8I5Jfyq_xnDHZEdFjtdBg
ships	https://pan.baidu.com/s/13khAoYjEFdHz49DV8U3usw

Training

• Download ImageNet pretrained vgg16_bn, and put it in pretrained folder. Or you can define your pre-trained model path in models/backbones/vgg.py

• To train our model on ShanghaiTech PartA, run

  sh train.sh
  

  • To train our model on CARPK, run

  sh train.sh (To modify the dataset in the train.sh)
  or
   CUDA_VISIBLE_DEVICES='0' python -m torch.distributed.launch --nproc_per_node=1 --master_port=10001 
    --use_env main.py 
    --lr=0.0001 
    --backbone="vgg16_bn" 
    --ce_loss_coef=1.0 
    --point_loss_coef=5.0 
    --eos_coef=0.5 
    --dec_layers=2 
    --hidden_dim=256 
    --dim_feedforward=512 
    --nheads=8 
    --dropout=0.0 
    --epochs=1500 
    --dataset_file="CARPK" 
    --eval_freq=5 
    --output_dir='pet_model'
  

Evaluation

• Modify eval.sh
  • change --resume to your local model path
• Run

sh eval.sh

Permission

This code is for academic purposes only. Contact: Yujia Liang (yjlianghust.edu.cn)

Acknowledgement

We thank the authors of PET for open-sourcing their work.