After labeling with the labeling tool, we would have an Image folder and a Label folder with same inner structure (i.e. for all .bmp file in Image, we have a .txt file with same relative path in Label). If you have multiple image folders and label folders, you should combine all image folders into one image folder and same for label folders, with same inner folder structure.
Run preprocess_imgs_rcnn.py will convert BMP images to JPG images and rename all files to number.txt or number.jpg directly under image or label folder.
The command is
python preprocess_imgs_rcnn.py --old_img_folder Image --old_label_folder Label --new_img_folder new_img --new_label_folder new_lbl --quality 95quality specifies the quality of the generated jpgs.
preprocess_imgs_rcnn_jpg.py does the same job except that it expects the input iamges to be jpgs.
The voc_dataset_build.py build a VOC2007 format data folder based on the above step, which includes XML building and coordinate shifting.
Suppose you want N test images and V validation images, you should use the following command:
python voc_dataset_build.py --voc_folder VOC2007 --img_folder new_img --label_folder new_lbl --test_num N --val_num VThe generated folder VOC2007 will be in VOC format.
If you happen to have images (in folder old_img) with higher magnification (e.g. 2x large) and want to test them, use the command below:
python image_rescale.py --input old_img --output new_img --scale 0.5The images inside new_img should be in correct scale as the network and you can pass them into the network.
Note: our network will perform worse on rescaled images.