SKU110K_CVPR19 shelf target detection and reproduction details (including independent image test)

Source code: https://github.com/eg4000/SKU110K_CVPR19
Paper address: https://arxiv.org/pdf/1904.00853.pdf

1. Preliminary preparation

1. Operating environment

Keras == 2.2.5
keras-retinanet == 0.5.1
tensorflow-gpu == 1.15.4
numpy == 1.19.2
opencv-python == 3.1.0.5
tqdm == 4.50.2
pandas == 0.23.4

pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt

2. Dataset

There is a data set download address in the source code, which contains image and annotation

2. Code modification

1. Copy files

Put the train.py file and class_mappings.csv in the main folder for easy operation

cp ./object_detector_retinanet/keras_retinanet/bin/train.py class_mappings.csv  ./

2. Modify the path where the dataset is located

Modify the root_dir function in the ./object_detector_retinanet/utils.py file

3. Add code

Add two lines to lines 24-25 in the ./object_detector_retinanet/utils/image.py file to avoid error reporting

from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True

3. Start training

1. train.py

Specification of relevant parameters

If not trained, then:

python -u ./train.py csv           

specified gpu

python -u ./train.py --gpu 1 csv           

If it has been trained, you can continue training with the most recently saved h5 model

python -u ./train.py --snapshot "SKU110K/snapshot/Thu_Feb_25_16_18_58_2021/resnet50_csv_03.h5" csv

Successful case:

2. train_iou.py

Also put train_iou.py into the main folder for easy training

cp ./object_detector_retinanet/keras_retinanet/bin/train_iou.py ./

start training iou

python -u ./train_iou.py --weights snapshot/Tue_May_26_08_17_14_2020/resnet50_csv_05.h5 csv

4. Start the test (test in the SKU dataset)

1. predict.py

Also put predict.py into the main folder

cp ./object_detector_retinanet/keras_retinanet_bin/predict.py to ./

2. Test

python -u ./predict.py csv  "./SKU110K/snapshot/Thu_May_28_02:36:24_2020/iou_resnet50_csv_01.h5" --hard_score_rate=0.5 

A certain result:

The folder where the result exists:

5. Test your own pictures (no csv)

1. Code modification

a. Copy predict.py and rename it to predict_test_demo.py

b. Copy predict_iou.py and rename it to predict_iou_test.py

c. Modify predict_test_demo.py

1. Comment out the reference to predict_iou and change it to predict_iou_test
   

2. Comment out all the parts containing args in the main function, because if you have args, you must enter data types such as csv/coco, and we only have the image path
   
   

3. model loaded directly
   

4. In the predict function, the generator is replaced with the image folder path, and the corresponding parameters are also modified
   

d. Modify predict_iou_test.py

1. Add related functions to be used later: preprocess_image, resize_image
   
2. Modify the predict function:
   modify the parameters, comment out the code related to csv, and add the code to read the pictures under the folder
   
3. Save
   and comment out the code related to csv, and comment out the code that draws the annotation frame
   

2. A picture of running your own tests

python -u ./predict_test_demo.py 

result:

The effect is good!

Reference blog: https://blog.csdn.net/qq_35975447/article/details/106349912