Open mzur opened 5 years ago
The new pipeline with PyTorch now uses the MMDetection random cropping, which should return smaller images unchanged. I haven't tried, though. Also, novelty detection will be an issue, as it also processes images in 512x512 windows.
The training dataset for Mask R-CNN is composed of 512x512px image crops. Smaller original images are not supported. Maybe change this to use the whole original image if it is smaller than 512x512px? As all training images need to have the same size, pad images with zeros to the largest image (crop) of the dataset (max 512x512).