PyTorch-YOLOv3

1. Installation

1.1. git clone

git clone https://github.com/yangzhaonan18/yolov3_trafficSign_pytorch

1.2. Download pretrained weights

1. Download detection(yolov3) weights: Baidu network disk link(234MB): https://pan.baidu.com/s/1BWySwi22nsFTB7-c0ualZA6

download and put it at: ./checkpoints/yolov3_ckpt_33.pth

2. Download classifier(CNN) weights: Baidu network disk link(43MB): https://pan.baidu.com/s/1Id65qVFrAp-S5G--57LG2Q6

download and put it at: ./ALL_sign_data/checkpoints/

1.3. Detect and classify an image

python3 detection_and_classification.py

detect the images in "./image_for_detect/Tinghua100K/", and the images with results will be saved in "./output", the "Tinghua100K_result.json" result will be saved in "./result/"

images with results:

2. How to test the results?

cd Tinghua100K_data/python/
python3 my_result_classes.py

results is:

iou:0.5, size:[0,400), types:[w55, ...total 53...], accuracy:0.8845589434208381, recall:0.9304519337964154
iou:0.5, size:[0,32), types:[w55, ...total 53...], accuracy:0.8160990712074303, recall:0.885752688172043
iou:0.5, size:[32,96), types:[w55, ...total 53...], accuracy:0.9297398348652438, recall:0.9740492900277461
iou:0.5, size:[96,400), types:[w55, ...total 53...], accuracy:0.9261477045908184, recall:0.8672897196261682
iou:0.5, size:[0,400), types:w55, accuracy:0.9611650485436893, recall:0.908256880733945
iou:0.5, size:[0,400), types:p27, accuracy:0.9156626506024096, recall:0.9047619047619048
iou:0.5, size:[0,400), types:il80, accuracy:0.9746192893401016, recall:0.9746192893401016
iou:0.5, size:[0,400), types:i1, accuracy:0.8571428571428571, recall:1.0
iou:0.5, size:[0,400), types:il100, accuracy:0.89, recall:0.967391304347826
......
......

3. How to train my dataset?

3.1. Train YOLOv3 detection

1. Download pretrained weights(on COCO): darknet53.conv.74

cd ./weights
bash download_weights.sh

2. Train YOLOv3 detection

cd ../
python3 train.py --data_config config/Tinghua100K.data --pretrained_weights weights/darknet53.conv.74

The YOLOv3 training weights will be saved in ./checkpoints/

3.2. Train CNN classifier

1. Download traffic sign data to train classifier Baidu network disk link: https://pan.baidu.com/s/133wOElvWHn0Fm4RzOGLk3w and unzip it in ALL_sign_data/ALL_data_in_2_train/

cd ./ALL_sign_data/
bash  run.sh

The train weights will be saved in ./ALL_sign_data/checkpoints

4. how to train your dataset for YOLOv3?

4.1 Prepare dataset

1. The data set is stored in folder : /headless/Desktop/yzn_file/DataSet_traffic_sign/ which contains 3 datasets:

• CCTSDB_changsha
• GTSDB
• Tinghua_100K

2. each file contains three files(most important)

• images_jpg : jpg images
• labels: txt label for YOLOv3
• labels_xml: xml label for VOC dataset

3. scripts

2_2voc_label.py : convert the three files "CCTSDB_changsha", "GTSDB", "Tinghua_100K" train and test label from .xml(VOC) to .txt(YOLOv3) and create the image path file of three datasets.

The "train.txt" and "test.txt" contain the images's absolute path will be saved in ./ALL_DATA

4.2. Creat the images's absolute path

cd  /headless/Desktop/yzn_file/DataSet_traffic_sign/
python3 2_2voc_label.py

4.3. Train

cd ../yolov3_trafficSign_pytorch


#  train from yolo pretrained weights:darknet53.conv.74
python3 train.py --pretrained_weights weights/darknet53.conv.74

or 
#  train from my pretrained weights:yolov3_ckpt_33.pth
python3 train.py --pretrained_weights checkpoints/yolov3_ckpt_33.pth

• Train logs

---- [Epoch 0/300, Batch 39/19133] ----
+------------+--------------+--------------+--------------+
| Metrics    | YOLO Layer 0 | YOLO Layer 1 | YOLO Layer 2 |
+------------+--------------+--------------+--------------+
| grid_size  | 38           | 76           | 304          |
| loss       | 20.597389    | 15.203256    | 22.110357    |
| x          | 0.153537     | 0.110538     | 0.065873     |
| y          | 0.059418     | 0.069747     | 0.058901     |
| w          | 2.156327     | 0.737079     | 0.113998     |
| h          | 1.517446     | 0.609778     | 0.385179     |
| conf       | 16.404194    | 13.595131    | 20.615255    |
| cls        | 0.306469     | 0.080982     | 0.871151     |
| cls_acc    | 100.00%      | 100.00%      | 100.00%      |
| recall50   | 0.000000     | 0.000000     | 0.000000     |
| recall75   | 0.000000     | 0.000000     | 0.000000     |
| precision  | 0.000000     | 0.000000     | 0.000000     |
| conf_obj   | 0.283539     | 0.095890     | 0.186076     |
| conf_noobj | 0.138449     | 0.103401     | 0.168427     |
+------------+--------------+--------------+--------------+
Total loss 57.91100311279297
---- ETA 3:00:59.369194
path ['/headless/Desktop/yzn_file/DataSet_traffic_sign/CCTSDB_changsha/images_jpg/train/06006.jpg']