Image Mixing and Deletion

• @naveedSurveyImageMixing2023
• [Cutout] and CutMix argues that hindering image regions enforces the classifier to learn from the partially visible objects and understand the overall structure
• CutMix verifies this argument by showing enhanced focus towards the target class in
• Opposite to this, MixUp has shown to improve classifier’s calibration and reduced prediction uncertainity
• mean of predictions vs accuracy where the confidence distribution for MixUp trained model is evenly distributed against the standard model whose disovertribution is towards higher conficence i.e. confidence
• Similarly, the loss contours obtained for a network trained with MixUp are smooth as compared to sharp contours in standarad training
• better generalization and robustness of MixUp against adversarial attacks.
• Mixup
• Cut and Delete
• Cutout
• Random Erasing
• Hide and Seek
• GridMask
• Adversarial Spatial Dropout for Occlusion
• Cut and Mix
• CutMix
• Attentive CutMix
• AttributeMix
• RICAP
• Mixed Example
• CowMask
• ResizeMix
• SaliencyMix
• Intra-Class Part Swapping
• SnapMix
• KeepAugment
• Visual Context Augmentation
• Cut, Paste and Learn
• Manifold MixUp
• AugMix
• SmoothMix
• Co-Mixup
• Sample Pairing
• Puzzle Mix
• ReMix