Learning to resize: Replace a front-end resizer in deep networks by a learnable non-linear resizer
Google Research

Deep computer vision models can benefit greatly from replacing a fixed linear resizer which you use to downsample Imagenet images before training with a well-designed, learned, nonlinear resizer.

Structure of the learned resizer is specific; not just adding more generic convolutional layers to the baseline model. Looks like it strives to encode some extra information in the downsampled image. From there stems the extra perfromance on Imagenet.

This work shows that a generically deeper model can be improved upon w/ a well-designed front-end, task-optimized, processor.

Looking ahead: probably there’s a lot of room for work on task-optimized pre-processing modules for computer vision and other tasks.

📝 Paper
No code yet

#cv #paper_explained