Multi-View Object Detection

💡Introduction

This is our submitted paper implemented using PyTorch.

🚩Dataset

/dataset/powerdata.yaml: Private Power Data 
    # Due to the signing of a confidentiality agreement, this dataset is not publicly available at this time.
/dataset/publicallpower.yaml: Public Power Data
    # CPLID: https://github.com/InsulatorData/InsulatorDataSet
    # IDID: https://ieee-dataport.org/competitions/insulator-defect-detection
    # VPMBGI: https://github.com/phd-benel/VPMBGI
/dataset/VisDrone.yaml:VisDrone2019-DET 
    # https://github.com/VisDrone/VisDrone-Dataset
other

🛠️Quick Start Examples

Install

export WANDB_MODE=disabled
conda activate DA #debug
conda activate ObjectDetection #train

# clone the project and configure the environment.
git clone https://github.com/LiuJiaji1999/MFCO.git
# the version of ultralytics is '8.3.9'           
# GPU - 24G NVIDIA GeForce RTX 3090 
# CPU - 12th Gen Intel(R) Core(TM) i9-12900
pip install -r OD-requirements.txt
attention:
  - opencv-python==4.9.0.80
  - opencv-python-headless==4.9.0.80
  - torch==1.9.0
  - torchvision==0.13.0+cu113
  - torchaudio==0.12.0+cu113

Train

python train.py 
# save outputlog
nohup python train.py > /log/XXX.log 2>&1 & tail -f /log/XXX.log

Test

Since github cannot push big file, we put '.pt' into the Google Drive,so you can directly test :
# https://drive.google.com/drive/folders/1SSlZEZvlthQqWaZFEPYCWwkEe-40fqsX
python val.py # test dataset 
python detect.py # Visualization of detect results
nohup python feature.py > /home/lenovo/data/liujiaji/powerGit/mvod/log/feature.log 2>&1 & tail -f /home/lenovo/data/liujiaji/powerGit/mvod/log/feature.log 

📢Plan

We will modify network to the enabling fair comparison with other MVOD methods.
    - ✅ MFFN-YOLO：https://github.com/zhanxn666/MFFN_YOLO (Camouflaged Object Detection)
    - 
And thanks again to the sprirt of the open-source:https://github.com/stars/LiuJiaji1999/lists/mvod

📝RT-DETR result

/rtdetr
model: rtdetr-l.yaml
datasets: Private Power Data, Public Power Data, VisDrone2019-DET
(640, 640)+ours: CUDA out of Memory

Method	pin-un	pin-ru	pin-de	insulator-bu	insulator-de	insulator-di	mAP(%)
(640,640)	22.3	67.3	48.0	71.5	69.0	29.9	51.3
(640,640)+ours	-	-	-	-	-	-	-
(384,384)	6.04	48.4	28.0	54.7	45.8	23.4	34.4
(384,384)+ours	10.6	57.4	33.1	66.0	66.6	35.2	44.8

Method	CPLID-defect	VPMBGI-defect	IDID-flashover	IDID-broken	mAP(%)
(640,640)	99.5	89.6	94.1	94.7	94.5
(640,640)+ours	-	-	-	-	-
(384,384)	97.9	80.1	81.0	82.5	85.4
(384,384)+ours	99.5	99.7	88.4	94.5	92.8

Method	ign-reg	ped	peo	bic	car	van	tru	tri	aw-tri	bus	mAP (%)
(640, 640)	38.9	28.0	14.9	79.1	41.3	46.2	28.5	18.7	56.8	42.7	39.5
(640,640)+ours	-	-	-	-	-	-	-	-	-	-	-
(384, 384)	19.3	14.9	6.0	66.4	29.7	32.0	14.6	11.2	44.6	23.9	26.3
(384,384)+ours	23.5	19.2	5.3	69.6	30.1	36.4	16.0	14.4	53.0	28.3	29.6

🌟Explanation of the file

1. main_profile.py ：model.info
2. test_yaml.py ：test all yaml is run 
3. heatmap.py ：heatmap
4. get_FPS.py ：compute model param、inference-time、FPS
5. get_model_erf.py ： erf
6. test_other.py ： debug
7. plot_result.py：visualize loss and metrics
8. plot_all.py:visualize loss and metrics in one picture
9. feature.py：visualize feature distributions

Personal Debug

print('一. trainer.py/get_dataset 先从yaml文件获取 train')
print('二. trainer.py/get_dataloader 开始加载训练数据')
print('三. detect/train.py/build_dataset 开始真正构建数据集')
print('四. bulid.py/build_yolo_dataset 构建YOLO数据集')
print('五. dataset.py/build_transforms 开始数据增强')
print('六. augment.py/v8_transforms 开始执行数据增强函数，') #随机增强方式直接替换原图送进模型    
print('七.ultralytics/data/base.py/get_image_and_label，数据增强后的图片-标签对应'）