Cut mix: Regularization strategy to train strong classifiers with localizable features
CutMix is a regularization technique that trains strong classifiers with localizable features. It involves randomly cropping a region from an image and replacing it with a patch from another image. The ground truth labels are also mixed proportionally to the pixel ratio. This allows for no information loss unlike dropout methods, while still enhancing localization ability by learning from partial views. CutMix outperforms other data augmentation techniques like Mixup by generating more natural composite images and helps models better localize objects. It has been shown to work well as a complementary technique to other regularizers and achieves state-of-the-art results on ImageNet and CIFAR-10 classification tasks when implemented in PyTorch.
Cut mix: Regularization strategy to train strong classifiers with localizable features
- 1. CutMix - Regularization Strategy to Train Strong Classifiers with Localizable Features Changjin Lee
- 2. Intro ● Many data augmentation and regularization methods have been proposed for vision tasks ● Random feature removal regularization - they work OK! ○ dropout ○ regional dropout - remove random spatial regions ● However, regional dropouts lead to information loss - severe conceptual limitation ❖ How to maximally utilize the deleted regions while preserving generalization and localization effects using regional dropout? ❖ The paper addresses this by CutMix - replaces the deleted region with a patch from another image.
- 3. CutMix Intro ● Crop a region and replace with a patch from another image ● The ground truth labels are also mixed proportionally to the pixel ratio of the two images Advantages ● No information loss ● Enhance localization ability - should learn to identify the object from a partial view
- 4. Comparison: Regional Dropout ● CutMix is similar to regional dropout in that they both crop a portion of image ● Regional Dropout - randomly remove a portion of image ● CutMix - randomly crop a portion and replace with a patch from another image Comparison: Synthesizing training data ● A synthesizing technique such as Stylizing ImageNet focuses more on shape than texture ● CutMix requires only trivial additional cost for training while generating new samples Comparison: Mixup ● Mixup samples introduce locally ambiguous and unnatural images
- 5. Complementary to other models ● CutMix is a great complementary to weight decay, batch normalization, and adding noises ● CutMix operates only on data level
- 6. CutMix Algorithm training images: (W*H*C) Binary Mask: (W*H) new training sample labels combination ratio: sampled from Beta(ɑ,ɑ) -> ɑ=1 1-λ λ B
- 7. CutMix on Class Activation Map (CAM) ● Vanilla ResNet-50 ● Cutout focuses on less discriminative parts like belly ● Mixup fully uses the pixels but it’s unnatural and confusing model which object to choose ● This confusion results in suboptimal performance ● CutMix successfully localize the two objects confused
- 8. Performances
- 10. Performance: CIFAR-10 Layer 0: input level (best) Layer 1: after conv-bn Layer 2: after layer 1 … Variations of CutMix lead to performance degradation -> Original is the best!
- 13. Dive Deeper… ❖ Random Croppings sometimes replace with useless images and this definitely reduces the performance ➢ Possible improvements: 1) Object Detection (inefficient?), 2) limiting cropped size i.e, lambda=U(0.5,0.8)? ❖ Weakly Supervised Object Localization ❖ Image Captioning Bad Good
- 14. References ● https://arxiv.org/pdf/1905.04899.pdf ● https://github.com/clovaai/CutMix-PyTorch Blog Post ● https://jasonlee-cp.github.io/paper/CutMix/