This is the implementation of "YOLOv2" for Object Detection.
Original paper: J. Redmon and A. Farhadi. YOLO9000: Better, Faster, Stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. link
Please build the source file according to the procedure.
$ mkdir build
$ cd build
$ cmake ..
$ make -j4
$ cd ..
- The PASCAL Visual Object Classes Challenge 2012 (VOC2012)
This is a set of images that has an annotation file giving a bounding box and object class label for each object in one of the twenty classes present in the image.
Link: official
Please create a link for the dataset.
The following hierarchical relationships are recommended.
datasets
|--Dataset1
| |--trainI
| | |--image1.png
| | |--image2.bmp
| | |--image3.jpg
| |
| |--trainO
| | |--label1.txt
| | |--label2.txt
| | |--label3.txt
| |
| |--validI
| |--validO
| |--testI
| |--testO
| |
| |--detect
| |--image4.png
| |--image5.bmp
| |--image6.jpg
|
|--Dataset2
|--Dataset3
- Class List
Please set the text file for class names.
$ vi list/VOC2012.txt
In case of "VOC2012", please set as follows.
aeroplane
bicycle
bird
boat
bottle
bus
car
cat
chair
cow
diningtable
dog
horse
motorbike
person
pottedplant
sheep
sofa
train
tvmonitor
- Input Image
You should substitute the path of training input data for "<training_input_path>", test input data for "<test_input_path>", detection input data for "<detect_path>", respectively.
The following is an example for "VOC2012".
$ cd datasets
$ mkdir VOC2012
$ cd VOC2012
$ ln -s <training_input_path> ./trainI
$ ln -s <test_input_path> ./testI
$ ln -s <detect_path> ./detect
$ cd ../..
- Output Label
You should get the id (class number), x-coordinate center, y-coordinate center, width, and height from the class and coordinate data of bounding boxes in the XML file and normalize them.
Please follow the steps below to convert XML file to text file.
Here, you should substitute the path of training XML data for "<training_xml_path>", test XML data for "<test_xml_path>", respectively.
The following is an example for "VOC2012".
$ ln -s <training_xml_path> ./datasets/VOC2012/trainX
$ ln -s <test_xml_path> ./datasets/VOC2012/testX
Please create a text file for training data.
$ vi ../../scripts/xml2txt.sh
You should substitute the path of training XML data for "--input_dir", training text data for "--output_dir", class name list for "--class_list", respectively.
#!/bin/bash
SCRIPT_DIR=$(cd $(dirname $0); pwd)
python3 ${SCRIPT_DIR}/xml2txt.py \
--input_dir "datasets/VOC2012/trainX" \
--output_dir "datasets/VOC2012/trainO" \
--class_list "list/VOC2012.txt"
The data will be converted by the following procedure.
$ sh ../../scripts/xml2txt.sh
Please create a text file for test data.
$ vi ../../scripts/xml2txt.sh
You should substitute the path of test XML data for "--input_dir", test text data for "--output_dir", class name list for "--class_list", respectively.
#!/bin/bash
SCRIPT_DIR=$(cd $(dirname $0); pwd)
python3 ${SCRIPT_DIR}/xml2txt.py \
--input_dir "datasets/VOC2012/testX" \
--output_dir "datasets/VOC2012/testO" \
--class_list "list/VOC2012.txt"
The data will be converted by the following procedure.
$ sh ../../scripts/xml2txt.sh
Please set the shell for executable file.
$ vi scripts/train.sh
The following is an example of the training phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.
#!/bin/bash
DATA='VOC2012'
./YOLOv2 \
--train true \
--augmentation true \
--epochs 300 \
--dataset ${DATA} \
--class_list "list/${DATA}.txt" \
--class_num 20 \
--size 608 \
--batch_size 8 \
--prob_thresh 0.03 \
--Lambda_noobject 0.1 \
--lr_init 1e-5 \
--lr_base 1e-4 \
--lr_decay1 1e-5 \
--lr_decay2 1e-6 \
--gpu_id 0 \
--nc 3
Please execute the following to start the program.
$ sh scripts/train.sh
Please set the shell for executable file.
$ vi scripts/test.sh
The following is an example of the test phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.
#!/bin/bash
DATA='VOC2012'
./YOLOv2 \
--test true \
--dataset ${DATA} \
--class_list "list/${DATA}.txt" \
--class_num 20 \
--size 608 \
--prob_thresh 0.03 \
--Lambda_noobject 0.1 \
--gpu_id 0 \
--nc 3
Please execute the following to start the program.
$ sh scripts/test.sh
Please set the shell for executable file.
$ vi scripts/detect.sh
The following is an example of the detection phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.
#!/bin/bash
DATA='VOC2012'
./YOLOv2 \
--detect true \
--dataset ${DATA} \
--class_list "list/${DATA}.txt" \
--class_num 20 \
--size 608 \
--prob_thresh 0.03 \
--gpu_id 0 \
--nc 3
Please execute the following to start the program.
$ sh scripts/detect.sh
Please set the shell for executable file.
$ vi scripts/demo.sh
The following is an example of the demo phase.
If you want to view specific examples of command line arguments, please view "src/main.cpp" or add "--help" to the argument.
#!/bin/bash
DATA='VOC2012'
./YOLOv2 \
--demo true \
--cam_num 0 \
--dataset ${DATA} \
--class_list "list/${DATA}.txt" \
--class_num 20 \
--size 608 \
--prob_thresh 0.03 \
--gpu_id 0 \
--nc 3
Please execute the following to start the program.
$ sh scripts/demo.sh
This code is inspired by darknet, Yolo-v2-pytorch, yolov2.pytorch, yolo2-pytorch, furkanu/yolov2-pytorch, yxlijun/yolov2-pytorch and YOLOv2.
Loss function (default):
If the loss of term conf<noobj> is strong, "Not Detected" will occur frequently.
In the case, it is recommended to add --Lambda_noobject 0.1 to arguments, where the default value is 1.0.
If the initial value of learning rate is high, gradient values will diverge.
In the case, it is recommended to add --lr_init 1e-5, --lr_base 1e-4, --lr_decay1 1e-5 and --lr_decay2 1e-6 to arguments. (i.e., slow updation and fluctuation)
Transformation of 8 components:
- Flipping :
--flip_rate 0.5 - Scaling (i.e., Resize) :
--scale_rate 0.5 - Blurring (i.e., Applying an averaging filter) :
--blur_rate 0.5 - Change Brightness (i.e., Value in HSV) :
--brightness_rate 0.5 - Change Hue :
--hue_rate 0.5 - Change Saturation :
--saturation_rate 0.5 - Shifting :
--shift_rate 0.5 - Cropping :
--crop_rate 0.5
Please write the occurrence probability in each argument rate.
Here, 1.0 means that it always occurs.
If the distributions of training and test samples are quite similar, data augmentation has undesirable effect.
In the case, it is recommended to add --augmentation false to arguments, where the default value is true.
Anchors are useful for stable detection.
Please refer to cfg/anchor.txt to change the config.
0.57273 0.667385 # (1) Prior-width Prior-height
1.87446 2.06253 # (2) Prior-width Prior-height
3.33843 5.47434 # (3) Prior-width Prior-height
7.88282 3.52778 # (4) Prior-width Prior-height
9.77052 9.16828 # (5) Prior-width Prior-height
Please write the prior-size, where each grid size is 1.0.
If you want to change types of anchor, please write types on --na to arguments.
Multi-Scale Training allows the predictor to detect objects at various resolutions.
Please refer to cfg/resize.txt to change the config.
10 # types of image size to resize (i.e., it must match the number of lines for size in the text file.)
10 # iterations to switch
320 320 # (1) Width Height
352 352 # (2) Width Height
384 384 # (3) Width Height
416 416 # (4) Width Height
448 448 # (5) Width Height
480 480 # (6) Width Height
512 512 # (7) Width Height
544 544 # (8) Width Height
576 576 # (9) Width Height
608 608 # (10) Width Height
Detection performance is determined by the prediction result and two threshold.
Two threshold:
- Simultaneous probability with confidence and class score :
--prob_thresh 0.1 - IoU between bounding boxes in Non-Maximum Suppression :
--nms_thresh 0.5
If you allow over-detection, please decrease prob_thresh and increase nms_thresh. (e.g., --prob_thresh 0.05, --nms_thresh 0.75)