search

ZJULearning / pixel_link

Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018
MIT License
768 stars 254 forks source link

image size #51

Open mn7142 opened 6 years ago

mn7142 commented 6 years ago

size of images in icdar 2015 is 720*1280 and fix for all images in my dataset size of image is variant , is this important? Should the size of the images be fixed؟

ciciphus commented 6 years ago

the preprocessing stuff resize the images. Just set the training image size in the scripts

mn7142 commented 6 years ago

@ciciphus Where should I set the size؟ please help me to find this part of code

ciciphus commented 6 years ago

not sure what size u would like to set. In train.sh you can set the training image size(input of the neural network) , but that's not relate to your image size. Cuz it has some other operations like rotate and crop in preprocessing/ssd_vgg_preprocessing.py .

northeastsquare commented 5 years ago

tf_image.py resize_image函数里面可以这样保持比例： image = tf.image.resize_image_with_pad(image, size[0], size[1]) 在test阶段： def preprocess_for_eval(image, labels, bboxes, xs, ys, out_shape, data_format='NHWC', resize=Resize.WARP_RESIZE, do_resize = False, scope='ssd_preprocessing_train')