Support RR360 (Rotated Rectangle 360) Detection in TRR360D #731

vansin · 2023-02-15T13:53:56Z

1. Motivation

传统的 mmrotate的 oc le90 le135 定义的算法和数据集检测范围不支持 360的旋转模板检测，不能区分矩形的ABCD点的语义，所以本工作，定义了 r360 角度定义法，同时制作了DOTA格式的TRR360D 数据集，在此数据集上验证改造算法的有效性。

2. Modification

A new generated dataset （TRR360D ）
R360 angle define, q2rbox r2qbox funcition
RR360LocalVisualizer
DOTAR360Metric

3. 快速上手

mkdir RR360
cd RR360
export WORK_HOME=$(pwd)

安装前置依赖

conda create -n RR360 python=3.8 -y
conda activate RR360

pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html

pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0rc1"

cd $WORK_HOME
git clone https://github.com/open-mmlab/mmdetection -b dev-3.x
cd mmdetection 
pip install -v -e .

# 假设已经安装mmengine、mmcv 2.x、mmdetection
cd $WORK_HOME
git clone https://github.com/vansin/mmrotate -b RR360
cd mmrotate 
export MMROTATE_HOME=$(pwd)
pip install -v -e .

数据集下载

cd $MMROTATE_HOME
mkdir -p data/TRR360D
# git clone https://github.com/vansin/TRR360D.git data/TRR360D
wget https://openmmlab.vansin.top/datasets/ICDAR2019_MTD_HOQ.zip -O data/TRR360D/TRR360D.zip
cd $MMROTATE_HOME/data/TRR360D
unzip TRR360D.zip

600 个训练图片，240个测试图片

通过 browse_dataset.py 检测数据集的正确性

cd $MMROTATE_HOME
 python projects/RR360/tools/browse_dataset.py \
 projects/RR360/configs360/rotated_retinanet/rotated-retinanet-rbox-r360_r50_fpn_6x_ic19.py \
    --output-dir show_ann \
    # --show

数据集有由语义信息的，ABCD点是表格左上角开始，依次顺时钟的四个点。

configs360

推理已训练好的模型

cd $MMROTATE_HOME

wget https://openmmlab.vansin.top/work_dirs/RR360/configs360/rotated_rtmdet_x3_r/rotated_rtmdet_s_l1-3x-ic19_pt/epoch_36.pth

python projects/RR360/tools/test.py \
    projects/RR360/configs360/rotated_rtmdet_x3_r/rotated_rtmdet_s_l1-3x-ic19_pt.py \
    epoch_36.pth \
    --show-dir predict_result
    # --show

在 work_dirs/rotated_rtmdet_s_l1-3x-ic19_pt/{时间戳}/predict_result 可以看到训练好的模型推理结果

训练

cd $MMROTATE_HOME
python projects/RR360/tools/train.py \
    projects/RR360/configs360/rotated_rtmdet_x3_r/rotated_rtmdet_s_l1-3x-ic19_pt.py \
    --d

测试自行训练的

cd $MMROTATE_HOME
python projects/RR360/tools/test.py \
    work_dirs/RR360/configs360/rotated_rtmdet_x3_r/rotated_rtmdet_s_l1-3x-ic19_pt/rotated_rtmdet_s_l1-3x-ic19_pt.py \
    work_dirs/RR360/configs360/rotated_rtmdet_x3_r/rotated_rtmdet_s_l1-3x-ic19_pt/epoch_36.pth

5. Checklist

zytx121

LGTM

zytx121 · 2023-03-12T14:48:03Z

@RangiLyu hope merge as soon as possible

codecov · 2023-03-12T14:58:45Z

Codecov Report

❗ No coverage uploaded for pull request base (dev-1.x@35c766e). Click here to learn what that means.
Patch has no changes to coverable lines.

❗ Current head b7088b6 differs from pull request most recent head f2e43c3. Consider uploading reports for the commit f2e43c3 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             dev-1.x     #731   +/-   ##
==========================================
  Coverage           ?   79.35%           
==========================================
  Files              ?      100           
  Lines              ?     7274           
  Branches           ?     1040           
==========================================
  Hits               ?     5772           
  Misses             ?     1164           
  Partials           ?      338

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

projects/RR360/readme.md

projects/RR360/tools/test.py

vansin · 2023-03-13T03:28:57Z

@RangiLyu 加TODO 的意思是要readme 写一下吗？

RangiLyu · 2023-03-13T03:57:39Z

@RangiLyu 加TODO 的意思是要readme 写一下吗？

代码里加上 TODO:xxxxx 的注释说明一下这么做是暂时的，之后会修改

RangiLyu · 2023-03-13T03:58:13Z

@RangiLyu 加TODO 的意思是要readme 写一下吗？

Readme最好也说明一下情况

vansin · 2023-03-13T04:03:42Z

@RangiLyu 加TODO 的意思是要readme 写一下吗？

Readme最好也说明一下情况

ok

lzh420202 · 2023-05-04T03:51:25Z

Does this coding directly apply to existing models to improve or degrade performance? I am also concerned about how to accurately evaluate this method, since the actual orientation information is not effectively represented in the mAP.

vansin · 2023-05-04T08:48:08Z

Does this coding directly apply to existing models to improve or degrade performance? I am also concerned about how to accurately evaluate this method, since the actual orientation information is not effectively represented in the mAP.

please see~

mmrotate/projects/RR360/evaluation/functional/mean_ap.py

Line 95 in 65a9a2c

    
           i] >= iou_thr and distance_gt_pred_angle < angle_thr * np.pi / 180:  # noqa: E501

https://github.com/open-mmlab/mmrotate/blob/dev-1.x/projects/RR360/evaluation/metrics/dota_r360_metric.py

dota_r360_metric

lzh420202 · 2023-05-04T09:11:01Z

Does this coding directly apply to existing models to improve or degrade performance? I am also concerned about how to accurately evaluate this method, since the actual orientation information is not effectively represented in the mAP.

please see~

mmrotate/projects/RR360/evaluation/functional/mean_ap.py

Line 95 in 65a9a2c

i] >= iou_thr and distance_gt_pred_angle < angle_thr * np.pi / 180: # noqa: E501

https://github.com/open-mmlab/mmrotate/blob/dev-1.x/projects/RR360/evaluation/metrics/dota_r360_metric.py

dota_r360_metric

Thanks, I understand the evaluation method, but did you test it on DOTA1.0? I just want to know if it has any advantages over the le90, le135, and oc on RS.
Thanks for your attention.

vansin · 2023-05-04T09:17:05Z

Does this coding directly apply to existing models to improve or degrade performance? I am also concerned about how to accurately evaluate this method, since the actual orientation information is not effectively represented in the mAP.

please see~

mmrotate/projects/RR360/evaluation/functional/mean_ap.py

Line 95 in 65a9a2c

i] >= iou_thr and distance_gt_pred_angle < angle_thr * np.pi / 180: # noqa: E501

https://github.com/open-mmlab/mmrotate/blob/dev-1.x/projects/RR360/evaluation/metrics/dota_r360_metric.py
dota_r360_metric

Thanks, I understand the evaluation method, but did you test it on DOTA1.0? I just want to know if it has any advantages over the le90, le135, and oc on RS. Thanks for your attention.

not test in DOTA 1.0, define a new metric named r360, the dataset satisfy follow condition can use this method

https://arxiv.org/pdf/2303.01894.pdf

welcome add my wechat van-sin

vansin added 30 commits February 5, 2023 19:45

init headet

94ac018

feat: add train and test

ae3e648

feat:update readme

e471b4e

feat: add h180

852f7b8

docs: update readme

b8f291e

feat: update readme

0ac0acb

feat: add datatsets

1b99811

docs: update h180 comments

5d100a9

feat: add configs-ic19-obb-obb configs

49b0ab5

feat: update metric

d949db2

feat: update metric

01a6735

feat: delete unnecessary

519f591

feat: init rtmdet

b511063

feat: update datasets

1dc74ab

feat: update local_visualizer comments

62eb949

feat: add benchmark.py

9a0aabb

docs: update

26ef5fd

docs: update

c139b4c

feat: update angle h180 in rtmdet

476f49f

feat: add roi_trans

066f311

feat: add 180

34fa858

docs: update RotatedBoxes

ff6cb00

feat: add r_l1_loss

94f618d

feat: add rl1 loss config

a5eaeb0

feat: update checkpoint hook

ed67e58

feat: change the weight

34e47cd

feat: update the loss weight 5

c855cf6

fix: reletive import

7b9f265

feat: fix the APH metric

39b8116

fix: fix the regularize_boxes

440f8a3

vansin added 7 commits March 12, 2023 21:09

docs: update

53102cf

fix: ICDAR2019_MTD_HOQ to TRR360D

e520ebf

fix: delete the unuse a l1 loss

2e1bf41

fix: delete the unuse model import

ad5e90c

docs: update

b690d79

fix: disable wandb

ef1e485

feat: add without l1_loss

b7088b6

zytx121 previously approved these changes Mar 12, 2023

View reviewed changes

zytx121 requested a review from RangiLyu March 12, 2023 14:46

zytx121 closed this Mar 12, 2023

zytx121 reopened this Mar 12, 2023

RangiLyu reviewed Mar 13, 2023

View reviewed changes

projects/RR360/readme.md Outdated Show resolved Hide resolved

docs: update en docs

d3151d5

vansin dismissed zytx121’s stale review via d3151d5 March 13, 2023 02:40

docs: update the datasets link

44dc690

RangiLyu reviewed Mar 13, 2023

View reviewed changes

projects/RR360/tools/test.py Show resolved Hide resolved

docs: add todo

f2e43c3

RangiLyu approved these changes Mar 13, 2023

View reviewed changes

RangiLyu merged commit 607db84 into open-mmlab:dev-1.x Mar 13, 2023

vansin changed the title ~~[WIP] Support RR360 (Rotated Rectangle 360) Detection in TRR360D~~ Support RR360 (Rotated Rectangle 360) Detection in TRR360D Mar 13, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support RR360 (Rotated Rectangle 360) Detection in TRR360D #731

Support RR360 (Rotated Rectangle 360) Detection in TRR360D #731

vansin commented Feb 15, 2023 •

edited

Loading

zytx121 left a comment

zytx121 commented Mar 12, 2023

codecov bot commented Mar 12, 2023 •

edited

Loading

vansin commented Mar 13, 2023

RangiLyu commented Mar 13, 2023

RangiLyu commented Mar 13, 2023

vansin commented Mar 13, 2023

lzh420202 commented May 4, 2023

vansin commented May 4, 2023 •

edited

Loading

lzh420202 commented May 4, 2023

vansin commented May 4, 2023 •

edited

Loading

Support RR360 (Rotated Rectangle 360) Detection in TRR360D #731

Support RR360 (Rotated Rectangle 360) Detection in TRR360D #731

Conversation

vansin commented Feb 15, 2023 • edited Loading

1. Motivation

2. Modification

3. 快速上手

configs360

5. Checklist

zytx121 left a comment

Choose a reason for hiding this comment

zytx121 commented Mar 12, 2023

codecov bot commented Mar 12, 2023 • edited Loading

Codecov Report

vansin commented Mar 13, 2023

RangiLyu commented Mar 13, 2023

RangiLyu commented Mar 13, 2023

vansin commented Mar 13, 2023

lzh420202 commented May 4, 2023

vansin commented May 4, 2023 • edited Loading

lzh420202 commented May 4, 2023

vansin commented May 4, 2023 • edited Loading

vansin commented Feb 15, 2023 •

edited

Loading

codecov bot commented Mar 12, 2023 •

edited

Loading

vansin commented May 4, 2023 •

edited

Loading

vansin commented May 4, 2023 •

edited

Loading