Learning Remote Sensing Object Detection with Single Point Supervision

This is the PyTorch implementation of the method in our paper "Learning Remote Sensing Object Detection with Single Point Supervision".

Preparation:

1. Requirement:

• [mmdetection 2.22.0]
• [DOTA_devkit]
• [iSAID_devkit]
• [dijkstra]

2. Generating data with single point labels:

• First, we use the DOTA_devkit and iSAID_devkit toolbox to generate cropped images with json annotations.
• Second, we add single point annotations in the json file of iSAID dataset data_process/iSAID_generate_single_point_annotation.py
• Third, we utilize the generated of iSAID json file to add single point information in the DOTA json file. data_process/DOTA_generate_single_point_annotation.py

Model training and validation:

1. Training PLUG:

• Run train.py with config configs_single_point/PLUG_r50_DOTA_512.py to train PLUG.

2. Referencing and validating PLUG:

• Run test_PLUG.py to reference PLUG and generate pseudo boxes of training data.
• Run data_process/calculate_mIoU.py to output the mIoU results of PLUG.

3. Training Faster-RCNN or Mask-RCNN:

• Run data_process/bbox2json.py to generate json of training data with pseudo boxes.
• Run train.py with config configs/faster_rcnn/faster_rcnn_r50_fpn_1x_dota.py to train Faster-RCNN.
• Run data_process/segm2json.py to generate json of training data with pseudo boxes and pseudo masks.
• Run train.py with config configs/mask_rcnn/mask_rcnn_r50_fpn_1x_dota.py to train Mask-RCNN.

4. Testing Faster-RCNN or Mask-RCNN:

• Run test.py to reference and evaluate Faster-RCNN and Mask-RCNN.
• Run DetVisGUI\DetVisGUI.py to visualize the detection results of different detectors conveniently. ( [DetVisGUI])

5. Other methods:

• We retrain P2BNet, WSDDN and OICR in our code based on [P2BNet] and [WSOD2].
• Run train.py with different configs to train the above methods.

6. Other codes:

• We split the training dataset according the object numbers in images to evaluate the effects of dense obejects. data_process/split_DOTA_image_and_json.py
• Run test_num.py to generate the pseudo boxes of sub dataset with different object numbers cyclically.
• Run data_process/bar_chart.py to generate the mIoU distribution of images with different object numbers.

Our model and data annotations:

• Please download from the [checkpoints(提取码：eh62)].

Citiation

If you find this work helpful, please consider citing:

@Article{PLUG-Det,
    author    = {He, Shitian and Zou, Huanxin and Wang, Yingqian and Li, Boyang and Cao, Xu and Jing, Ning},
    title     = {Learning Remote Sensing Object Detection with Single Point Supervision},
    journal   = {IEEE TGRS}, 
    year      = {2023},   
}

Contact

Welcome to raise issues or email to [email protected] for any question regarding this work.