README_en.md 8.2 KB

English | 简体中文

Rotated Object Detection

Table of Contents

Introduction

Rotated object detection is used to detect rectangular bounding boxes with angle information, that is, the long and short sides of the rectangular bounding box are no longer parallel to the image coordinate axes. Oriented bounding boxes generally contain less background information than horizontal bounding boxes. Rotated object detection is often used in remote sensing scenarios.

Model Zoo

Model mAP Lr Scheduler Angle Aug GPU Number images/GPU download config
S2ANet 73.84 2x le135 - 4 2 model config
FCOSR 76.62 3x oc RR 4 4 model config
PP-YOLOE-R-s 73.82 3x oc RR 4 2 model config
PP-YOLOE-R-s 79.42 3x oc MS+RR 4 2 model config
PP-YOLOE-R-m 77.64 3x oc RR 4 2 model config
PP-YOLOE-R-m 79.71 3x oc MS+RR 4 2 model config
PP-YOLOE-R-l 78.14 3x oc RR 4 2 model config
PP-YOLOE-R-l 80.02 3x oc MS+RR 4 2 model config
PP-YOLOE-R-x 78.28 3x oc RR 4 2 model config
PP-YOLOE-R-x 80.73 3x oc MS+RR 4 2 model config

Notes:

  • if GPU number or mini-batch size is changed, learning rate should be adjusted according to the formula lrnew = lrdefault * (batch_sizenew * GPU_numbernew) / (batch_sizedefault * GPU_numberdefault).
  • Models in model zoo is trained and tested with single scale by default. If MS is indicated in the data augmentation column, it means that multi-scale training and multi-scale testing are used. If RR is indicated in the data augmentation column, it means that RandomRotate data augmentation is used for training.

Data Preparation

DOTA Dataset preparation

The DOTA dataset is a large-scale remote sensing image dataset containing annotations of oriented and horizontal bounding boxes. The dataset can be download from Official Website of DOTA Dataset. When the dataset is decompressed, its directory structure is shown as follows.

${DOTA_ROOT}
├── test
│   └── images
├── train
│   ├── images
│   └── labelTxt
└── val
    ├── images
    └── labelTxt

For labeled data, each image corresponds to a txt file with the same name, and each row in the txt file represent a rotated bouding box. The format is as follows:

x1 y1 x2 y2 x3 y3 x4 y4 class_name difficult

Slicing data with single scale

The image resolution of DOTA dataset is relatively high, so we usually slice the images before training and testing. To slice the images with a single scale, you can use the command below

# slicing labeled data
python configs/rotate/tools/prepare_data.py \
    --input_dirs ${DOTA_ROOT}/train/ ${DOTA_ROOT}/val/ \
    --output_dir ${OUTPUT_DIR}/trainval1024/ \
    --coco_json_file DOTA_trainval1024.json \
    --subsize 1024 \
    --gap 200 \
    --rates 1.0
# slicing unlabeled data by setting --image_only
python configs/rotate/tools/prepare_data.py \
    --input_dirs ${DOTA_ROOT}/test/ \
    --output_dir ${OUTPUT_DIR}/test1024/ \
    --coco_json_file DOTA_test1024.json \
    --subsize 1024 \
    --gap 200 \
    --rates 1.0 \
    --image_only

Slicing data with multi scale

To slice the images with multiple scales, you can use the command below

# slicing labeled data
python configs/rotate/tools/prepare_data.py \
    --input_dirs ${DOTA_ROOT}/train/ ${DOTA_ROOT}/val/ \
    --output_dir ${OUTPUT_DIR}/trainval/ \
    --coco_json_file DOTA_trainval1024.json \
    --subsize 1024 \
    --gap 500 \
    --rates 0.5 1.0 1.5
# slicing unlabeled data by setting --image_only
python configs/rotate/tools/prepare_data.py \
    --input_dirs ${DOTA_ROOT}/test/ \
    --output_dir ${OUTPUT_DIR}/test1024/ \
    --coco_json_file DOTA_test1024.json \
    --subsize 1024 \
    --gap 500 \
    --rates 0.5 1.0 1.5 \
    --image_only

Custom Dataset

Rotated object detction uses the standard COCO data format, and you can convert your dataset to COCO format to train the model. The annotations of standard COCO format contains the following information

'annotations': [
    {
        'id': 2083, 'category_id': 9, 'image_id': 9008,
        'bbox': [x, y, w, h], # horizontal bouding box
        'segmentation': [[x1, y1, x2, y2, x3, y3, x4, y4]], # rotated bounding box
        ...
    }
    ...
]

It should be noted that bbox is the horizontal bouding box, and segmentation is four points of rotated bounding box (clockwise or counterclockwise). The bbox can be empty when training rotated object detector, and it is recommended to generate bbox according to segmentation. In PaddleDetection 2.4 and earlier versions, bbox represents the rotated bounding box [x, y, w, h, angle] and segmentation is empty. But this format is no longer supported after PaddleDetection 2.5, please download the latest dataset or convert to standard COCO format.

Installation

Models of rotated object detection depend on external operators for training, evaluation, etc. In Linux environment, you can execute the following command to compile and install.

cd ppdet/ext_op
python setup.py install

In Windows environment, perform the following steps to install it:

(1)Visual Studio (version required >= Visual Studio 2015 Update3);

(2)Go to Start --> Visual Studio 2017 --> X64 native Tools command prompt for VS 2017;

(3)Setting Environment Variables:set DISTUTILS_USE_SDK=1

(4)Enter ppdet/ext_op directory,use python setup.py install to install。

After the installation, you can execute the unittest of ppdet/ext_op/unittest to verify whether the external oprators is installed correctly.