Mix & Match: training convnets with mixed image sizes for improved accuracy, speed and scale resiliency.
Elad HofferBerry WeinsteinItay HubaraTal Ben-NunTorsten HoeflerDaniel SoudryPublished in: CoRR (2019)
Keyphrases
- improved accuracy
- input image
- image analysis
- image content
- image representation
- multiscale
- image collections
- image structure
- feed forward
- false matches
- image data
- image segmentation
- image retrieval
- low level
- image classification
- artificial neural networks
- template matching
- segmentation method
- test images
- training set
- spatial information
- image pixels
- multiple scales
- prediction accuracy
- single image
- image transformations
- image features
- feature points
- machine learning
- zero crossing
- similarity measure
- integral image
- edge detection
- image matching
- hough transform
- image regions
- keypoints
- segmentation algorithm
- natural images
- image quality
- high speed