Skip to content

Latest commit

 

History

History
167 lines (137 loc) · 4.42 KB

README.en.md

File metadata and controls

167 lines (137 loc) · 4.42 KB

YOLOV3-SUPER

Introduction

This is a YOLOV3 baseline written in PyTorch. The eval dataset used is PASCAL VOC(not use difficulty). The eval tool is the cocoAPI. The mAP gains the score as same as the original paper. Subsequently, we will continue to update the code by adding some new and efficient methods to make it more concise, lighter, faster in vehicle detection.

Results

name Train Dataset Val Dataset mAP(mine) notes
YOLOV3-*-544 2007trainval + 2012trainval 2007test 0.835 +cosine lr + multi-scale + smooth_L1
YOLOV3-*-544 2007trainval + 2012trainval cocoAPI 0.821 +cosine lr + multi-scale + smooth_L1

FPS

Implement Backbone Input Size Batch Size Inference Time FPS
Paper Darknet53 320 1 22ms 45
Paper Darknet53 416 1 29ms 34
Paper Darknet53 608 1 51ms 19
Our Darknet53 320 1 23ms 43
Our Darknet53 416 1 27ms 36
Our Darknet53 608 1 40ms 29

Note :

  • YOLOV3-*-544 means test image size is 544. "*" means the multi-scale.
  • In the test, the nms threshold is 0.5 and the conf_score is 0.01.
  • Now only support the single gpu to train and test.

Environment

Note: This is just a reference, the environment doesn’t have to be exactly the same.

  • Nvidia GeForce RTX 1080 Ti
  • CUDA10.1
  • CUDNN7.6.0
  • ubuntu 18.04
# install packages
pip install -r requirements.txt

Brief

  • Data Augmentation (SSD Augmentation methods)
  • Step lr Schedule
  • Multi-scale Training (320 to 608)
  • cosine lr

Prepared work

1、Git clone YOLOV3 repository

git clone https://gitee.com/vehicle_1/YOLOV3_SUPER.git

2、Download dataset

2.1 PASCAL_VOC

  • Download Pascal VOC dataset : VOC data.zip Passwords: 7vfh
  • The organizational structure of this data will be presented below:
    |--voc_data
    |--|--VOCtest-2007
    |--|--|--VOCdevkit
    |--|--|--|--VOC2007
    |--|--VOCtrainval-2007
    |--|--|--VOCdevkit
    |--|--|--|--VOC2007
    |--|--VOCtrainval-2012
    |--|--|--VOCdevkit
    |--|--|--|--VOC2012

2.2 U-DETRAC

  • Download Pascal VOC dataset : U-DETRAC
  • The organizational structure of this data will be presented below:
    |--U-DETRA_dataset
    |--|--DETRAC-Test-Annotations-XML
    |--|--DETRAC-Train-Annotations-XML
    |--|--DETRAC-test-data
    |--|--DETRAC-train-data

Prepare to do

  • first
    we should first adjust the config params in ./train/Detrac_data_preprocess/params_init_Detrac or ./train/Voc_data_preprocess/params_init_voc
    We should pay attention to the settings of these parameters:
    *"data_path" *"working_dir" *"DATA_PATH" *"PROJECT_PATH"
  • second
cd train/Detrac_data_preprocess
python Detrac_data_process.py

or

cd train/Voc_data_preprocess
python voc_data_process.py

then the dataset is ready

  • third
cd evaluate/evaluate_coco
python voc_convert2_coco.py --xml_dir ~~~

or

cd evaluate/evaluate_detrac_coco_api
python Detrac_convert2_coco.py

then the evaluation's gt_json for evaluate is ready

3、Download weight file

Make dir weight/ in the YOLOV3 and put the weight file in.

Train

Run the following command to start training and see the details in the darknet53/name/time_id/log.txt

cd train/
python train_model.py --config_name "VOC" --device_id 0 --config_model_name "yolov3"

Test

You should config evaluate/evaluate_coco/yolov3_config_voc_test.py or evaluate/evaluate_detrac_coco_api/yolov3_config_dtrac_test.py

cd evaluate/evaluate_coco/
python coco_eval_voc.py
cd evaluate/evaluate_detrac_coco_api/
python  coco_eval_detrac.py

The images can be seen in the evaluate/evaluate_coco/data/resultsorevaluate/evaluate_detrac_coco_api/data/results

TODO

  • model compression
  • focal loss
  • GIOU
  • Label smooth
  • Mixup
  • Mobilenet v1-v3
  • Mossaic
  • Poly-yolo
  • Gaussian-yolo
  • centernet-Gaussian kernel

Reference