Rotate-Yolov5

This repository is based on Ultralytics/yolov5, with adjustments to enable rotate prediction boxes.

Section I. Description

The codes are based on Ultralytics/yolov5, and several functions are added and modified to enable rotate prediction boxes.

The modifications compared with Ultralytics/yolov5 and their brief descriptions are summarized below:

data/rotate_ucas.yaml : Exemplar UCAS-AOD dataset to test the effects of rotate boxes
data/images/UCAS-AOD : For the inference of rotate-yolov5s-ucas.pt
models/common.py :

3.1. class Rotate_NMS : Non-Maximum Suppression (NMS) module for Rotate Boxes

3.2. class Rotate_AutoShape : Rotate Version of Original AutoShape, input-robust polygon model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and Rotate_NMS

3.3. class Rotate_Detections : Rotate detections class for Rotate-YOLOv5 inference results
models/rotate_yolov5s_ucas.yaml : Configuration file of rotate yolov5s for exemplar UCAS-AOD dataset
models/yolo.py :

5.1. class Rotate_Detect : Detect head for rotate-yolov5 models with rotate box prediction

5.2. class Rotate_Model : Rotate yolov5 models with rotate box prediction
utils/iou_cuda : CUDA extension for iou computation of polygon boxes

6.1. extensions.cpp : CUDA extension file

6.2. inter_union_cuda.cu : CUDA code for computing iou of polygon boxes

6.3. setup.py : for building CUDA extensions module polygon_inter_union_cuda, with two functions polygon_inter_union_cuda and polygon_b_inter_union_cuda
utils/autoanchor.py :

7.1. def rotate_check_anchors : Rotate version of original check_anchors

7.2. def rotate_kmean_anchors : Create kmeans-evolved anchors from rotate-enabled training dataset
utils/datasets.py :

8.1. def polygon_random_perspective : Data augmentation for datasets with polygon boxes (augmentation effects: HSV-Hue, HSV-Saturation, HSV-Value, rotation, translation, scale, shear, perspective, flip up-down, flip left-right, mosaic, mixup)

8.2. def polygon_box_candidates : Polygon version of original box_candidates

8.3. def rotate_random_perspective : Data augmentation for datasets with rotate boxes (augmentation effects: HSV-Hue, HSV-Saturation, HSV-Value, rotation, translation, scale, shear, perspective, flip up-down, flip left-right, mosaic, mixup)

8.4. class Rotate_LoadImagesAndLabels : Rotate version of original LoadImagesAndLabels

8.5. def rotate_load_mosaic : Loads images in a 4-mosaic, with rotate boxes

8.6. def rotate_load_mosaic9 : Loads images in a 9-mosaic, with rotate boxes

8.7. def rotate_verify_image_label : Verify one image-label pair for rotate datasets

8.8. def create_dataloader : Has been modified to include rotate datasets

8.9. class Albumentations : For albumentation augmentation
utils/general.py :

9.1. def xyxyxyxyn2xyxyxyxy : Convert normalized xyxyxyxy or segments into pixel xyxyxyxy or segments

9.2. def polygon_segment2box : Convert 1 segment label to 1 polygon box label

9.3. def polygon_inter_union_cpu : iou computation (polygon) with cpu

9.4. def polygon_box_iou : Compute iou of polygon boxes via cpu or cuda

9.5. def polygon_b_inter_union_cpu : iou computation (polygon) with cpu

9.6. def polygon_bbox_iou : Compute iou of polygon boxes via cpu or cuda

9.7. def polygon_nms_kernel : Non maximum suppression kernel for polygon-enabled boxes

9.8. def order_corners : Return sorted corners

9.9. def xywhrm2xyxyxyxy : Convert rotate xywhrm into xyxyxyxy, suitable for both pixel-level or normalized

9.10. def xyxyxyxy2xywhrm : Convert xyxyxyxy into rotate xywhrm, suitable for both pixel-level and normalized

9.11. def xywhn2xywh : Convert normalized xywh into pixel xywh

9.12. def rotate_segments2boxes : Convert segment labels to rotate box labels, i.e. (xy1, xy2, ...) to rotated boxes (x, y, w, h, re, im)

9.13. def rotate_scale_coords : Rescale coords (x, y, w, h, re, im) from img1_shape to img0_shape

9.14. def rotate_box_iou : Compute iou of rotate boxes via cpu or cuda

9.15. def rotate_bbox_iou : Compute iou of rotated boxes for class Rotate_ComputeLoss in loss.py via cpu or cuda

9.16. def rotate_non_max_suppression : Runs Non-Maximum Suppression (NMS) on inference results for rotated boxes
utils/loss.py :

10.1. class Rotate_ComputeLoss : Compute loss for rotate boxes
utils/metrics.py :

11.1. class Rotate_ConfusionMatrix : Rotate version of original ConfusionMatrix
utils/plots.py :

12.1. def polygon_plot_one_box : Plot one polygon box on image

12.2. def polygon_plot_one_box_PIL : Plot one polygon box on image via PIL

12.3. def polygon_plot_images : Polygon version of original plot_images

12.4. def rotate_plot_one_box : Plot one rotate box on image

12.5. def rotate_plot_one_box_PIL : Plot one rotate box on image via PIL

12.6. def rotate_output_to_target : Convert model output format [x, y, w, h, re, im, conf, class_id] to target format [batch_id, class_id, x, y, w, h, re, im, conf]

12.7. def rotate_plot_images : Rotate version of original plot_images

12.8. def rotate_plot_test_txt : Rotate version of original plot_test_txt

12.9. def rotate_plot_targets_txt : Rotate version of original plot_targets_txt

12.10. def rotate_plot_labels : Rotate version of original plot_labels
rotate_train.py : For training rotate-yolov5 models
rotate_test.py : For testing rotate-yolov5 models
rotate_detect.py : For detecting rotate-yolov5 models
requirements.py : Added python model shapely

Section II. How Does Rotate Boxes Work? How Does Rotate Boxes Different from Polygon Boxes?

Comparisons between Rotate-Yolov5 and Polygon-Yolov5

2. Model Head of Rotate-Yolov5

3. Illustration of Box Loss of Rotated Boxes

Section III. Installation

For the CUDA extension to be successfully built without error, please use CUDA version >= 11.2. The codes have been verified in Ubuntu 16.04 with Tesla K80 GPU.

# The following codes install CUDA 11.2 from scratch on Ubuntu 16.04, if you have installed it, please ignore
# If you are using other versions of systems, please check https://tutorialforlinux.com/2019/12/01/how-to-add-cuda-repository-for-ubuntu-based-oses-2/
# Install Ubuntu kernel head
sudo apt install linux-headers-$(uname -r)

# Pinning CUDA repo
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-ubuntu1604.pin
sudo mv cuda-ubuntu1604.pin /etc/apt/preferences.d/cuda-repository-pin-600

# Add CUDA GPG key
sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub

# Setting up CUDA repo
sudo add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/ /"

# Refresh apt repositories
sudo apt update

# Installing CUDA 11.2
sudo apt install cuda-11-2 -y
sudo apt install cuda-toolkit-11-2 -y

# Setting up path
echo 'export PATH=/usr/local/cuda-11.2/bin${PATH:+:${PATH}}' >> $HOME/.bashrc
# You are done installing CUDA 11.2

# Check NVIDIA
nvidia-smi
# Update all apts
sudo apt-get update
sudo apt-get -y upgrade

# Begin installing python 3.7
curl -o ~/miniconda.sh -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod +x ~/miniconda.sh
./miniconda.sh -b
echo "PATH=~/miniconda3/bin:$PATH" >> ~/.bashrc 
source ~/.bashrc
conda install -y python=3.7
# You are done installing python

The following codes set you up with the Rotate Yolov5.

# clone git repo
git clone https://github.com/XinzeLee/RotateObjectDetection
cd RotateObjectDetection/rotate-yolov5
# install python package requirements
pip install -r requirements.txt
# install CUDA extensions
cd utils/iou_cuda
python setup.py install
# cd back to rotate-yolov5 folder
cd .. && cd ..

Section IV. Rotate-Tutorial 1: Deploy the Rotate Yolov5s

Try Rotate Yolov5s Model by Following Rotate-Tutorial 1

Inference

<div class="highlight highlight-source-shell position-relative">
  <pre>
  $ python rotate_detect.py --weights rotate-yolov5s-ucas.pt --img 1024 --conf 0.75 \
      --source data/images/UCAS-AOD --iou-thres 0.4 --hide-labels</pre>
  <p align="center">
  <img src="https://user-images.githubusercontent.com/87064748/125021658-ad83eb80-e0ad-11eb-9a61-7824cc09b4ba.png" width="500">
  </p></div>

Test

<div class="highlight highlight-source-shell position-relative">
  <pre>
  $ python rotate_test.py --weights rotate-yolov5s-ucas.pt --data rotate_ucas.yaml \
      --img 1024 --iou 0.65 --task val</pre>
  <p align="center">
  <img src="https://user-images.githubusercontent.com/87064748/129825006-26677706-5da5-4584-9f7b-7a370e42b6d4.png" width="500">
  </p></div>

Train

<div class="highlight highlight-source-shell position-relative">
  <pre>
  $ python rotate_train.py --weights rotate-yolov5s-ucas.pt --cfg rotate_yolov5s_ucas.yaml \
      --data rotate_ucas.yaml --hyp hyp.ucas.yaml --img-size 1024 \
      --epochs 3 --batch-size 12 --noautoanchor --rotate --cache</pre></div>

Performance

<div class="highlight highlight-source-shell position-relative">
  4.1. Confusion Matrix
    <br/>
    <p align="center">
    <img src="https://user-images.githubusercontent.com/87064748/129825814-24680b53-c58f-4fda-a596-92347ee92fed.png" width="500">
    </p>
  4.2. Precision Curve
    <br/>
    <p align="center">
    <img src="https://user-images.githubusercontent.com/87064748/129825827-ac4c4298-8604-468a-b36e-638bc70bda35.png" width="500">
    </p>
  4.3. Recall Curve
    <br/>
    <p align="center">
    <img src="https://user-images.githubusercontent.com/87064748/129825828-61c0f977-9db3-4014-ad10-d99acfbd78c9.png" width="500">
    </p>
  4.4. Precision-Recall Curve
    <br/>
    <p align="center">
    <img src="https://user-images.githubusercontent.com/87064748/129825867-9a14d870-76f5-4df0-a0ff-2f9451828ac1.png" width="500">
    </p>
  4.5. F1 Curve
    <br/>
    <p align="center">
    <img src="https://user-images.githubusercontent.com/87064748/129825875-6dd339fd-7561-4d0e-a9fa-d4a0892250f5.png" width="500">
    </p></div>