Note

I'm solving scale invariant. If you have a good paper, you can email me by [email protected]. Thanks!

Improve yolo_v3 with latest paper

updated

• Data augmentation(release)
  
• Multi-scale training(release)
  
• Focal loss(increase 2 mAP, release)
  
• Single-Shot Object Detection with Enriched Semantics(incrase 1 mAP, not release)
  
• Soft-NMS(drop 0.5 mAP, release)
  
• Group Normalization(didn't use it in project, release)
  
• Recently updated: Modified the assign method of positive and negative samples(increase 0.6 mAP, release)
  
• Recently updated: Multi-scale testing(increase 2 mAP, release)
  

to do

• Deformable convolutional networks
  
• Scale-Aware Trident Networks for Object Detection
• Understanding the Effective Receptive Field in Deep Convolutional Neural Networks
  

performance on VOC2007

(score_threshold=0.01, iou_threshold=0.45, test_input_size=544)
If you want to get a higher mAP, you can set the score threshold to 0.01.
If you want to apply it, you can set the score threshold to 0.2.

1. initial with yolov3-608.weights
   

   size	mAP
   544	88.91
   multi scale	90.52

2. initial with darknet53.weights
The same performance as Tencent's reimplementation

|    size     |  mAP  |
|    :--:     | :---: |
|    544      | 79.32 |
| multi scale | 81.89 |

Usage

1. clone YOLO_v3 repository

   git clone https://github.com/Stinky-Tofu/YOLO_v3.git

2. prepare data
   (1) download datasets
   Create a new folder named data in the directory where the YOLO_V3 folder is located, and then create a new folder named VOC in the data/.
   Download VOC 2012_trainval 、VOC 2007_trainval 、VOC 2007_test, and put datasets into data/VOC, name as 2012_trainval、2007_trainval、2007_test separately.
   The file structure is as follows:
   |--YOLO_V3
   |--data
   |--|--VOC
   |--|--|--2012_trainval
   |--|--|--2007_trainval
   |--|--|--2007_test
   

   (2) convert data format
   You should set DATASET_PATH in config.py to the path of the VOC dataset, for example: DATASET_PATH = '/home/xzh/doc/code/python_code/data/VOC',and then
   

   python voc_annotation.py

3. prepare initial weights
   Download YOLOv3-608.weights firstly, put the yolov3.weights into yolov3_to_tf/, and then

   cd yolov3_to_tf
   python3 convert_weights.py --weights_file=yolov3.weights --dara_format=NHWC --ckpt_file=./saved_model/yolov3_608_coco_pretrained.ckpt
   cd ..
   python rename.py

4. Train
   

   python train.py

5. Test
   Download weight file yolo_test.ckpt

   python test.py --gpu=0 --map_calc=True --weights_file=model_path.ckpt
   cd mAP
   python main.py -na -np

Train for custom dataset

1. Generate your own annotation file train_annotation.txt and test_annotation.txt, one row for one image.
   Row format: image_path bbox0 bbox1 ...
   Bbox format: xmin,ymin,xmax,ymax,class_id(no space), for example:
   

   /home/xzh/doc/code/python_code/data/VOC/2007_test/JPEGImages/000001.jpg 48,240,195,371,11 8,12,352,498,14

2. Put the train_annotation.txt and test_annotation.txt into YOLO_V3/data/.
   
3. Configure config.py for your dataset.
   
4. Start training.
   

   python train.py

Reference:

paper:

• YOLOv3: An Incremental Improvement
  
• Foca Loss for Dense Object Detection
  
• Group Normalization
  
• Single-Shot Object Detection with Enriched Semantics
  
• An Analysis of Scale Invariance in Object Detection - SNIP
  
• Deformable convolutional networks
  
• Scale-Aware Trident Networks for Object Detection
  
• Understanding the Effective Receptive Field in Deep Convolutional Neural Networks
  

mAP calculate: mean Average Precision

Requirements

• Python2.7.12
  
• Numpy1.14.5
  
• Tensorflow.1.8.0
  
• Opencv3.4.1