Initial with YOLOV3-608.weights

Performance

train dataset: VOC 2012 + VOC 2007
test dataset: VOC 2007
test size: 544
test code: mAP (not use 07 metric)
test score threshold: 0.01

model	mAP	delta	release
baseline	84.3	0.0	yes
data agumentation	85.8	+1.5	yes
multi scale train	86.3	+0.5	yes
focal loss	88.3	+2.0	yes
des	xxx	+1.0	no
group normalization	xxx	xxx	yes
soft nms	xxx	-0.6	yes
label smooth	88.6	+0.3	yes
cosine learning rate	88.6	0.0	yes
GIOU	88.8	+0.2	yes
multi scale test	90.7	+1.9	yes

Usage

1. clone YOLO_v3 repository

   git clone https://github.com/Stinky-Tofu/Stronger-yolo.git

2. prepare data
   (1) download datasets
   Create a new folder named data in the directory where the Stronger-yolo 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:
   |--stronger-yolo
   |--|--v1
   |--|--v2
   |--|--v3
   |--data
   |--|--VOC
   |--|--|--2012_trainval
   |--|--|--2007_trainval
   |--|--|--2007_test
   (2) convert data format
   You should alter DATASET_PATH in config.py, for example:
   DATASET_PATH = /home/wz/doc/code/python_code/data/VOC
   

   cd Stronger-yolo/v1
   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 --data_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 stronger-yolo-v1-test.ckpt
   If you want to get a higher mAP, you can set the score threshold to 0.01、use multi scale test、flip test.
   If you want to use it in actual projects, or if you want speed, you can set the score threshold to 0.2.
   

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

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
  
• Bag of Freebies for Training Object Detection Neural Networks
  
• Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression
  

Requirements

software

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

hardware

• 12G 1080Ti