When do resizing on gluoncv + .rec?


Hi, I’m adapting this tuto https://gluon-cv.mxnet.io/build/examples_detection/finetune_detection.html to a use-case of license plate recognition. My dataset consists of pictures of variable size (between around 300 px width to 1000 px width). High-quality images may contain several small license plates, hence I need to keep the model operating at a reasonably high resolution. My question is the following:
When do I adjust pic sizes? Before im2rec-ing the pictures? Or downstream in the data loading process? Is there a specific primitive in gluoncv responsible for this size alignment?


If you know already which kind of resolution you need, you can use im2rec with the flag --resize to adjust the size of images. The advantage would be that your input file would be already significantly smaller and this will speed up reading the file from disk. If you need high resolution images, cropping the image could be an alternative.