You are tuning a convolutional neural network that detects road-signs in dash-cam video. Evaluation shows a sharp drop in mAP whenever a sign is partly hidden by tree branches or the rear of another vehicle. You decide to add occlusion data augmentation. Which augmentation policy is most likely to raise robustness to partial occlusion without introducing a large distribution shift that hurts performance on fully visible signs?
Apply CutMix so that 50% of every training image is replaced by pixels from another randomly chosen image, regardless of object location.
Overlay an opaque rectangular mask that hides 30-60% of the sign area in every single training image.
Replace the entire bounding box of each sign with a uniform gray patch in 10% of the images.
In about 25% of mini-batches, paste realistic object cut-outs (cars, foliage) so they cover 10-30% of randomly selected signs.
Pasting realistic object cut-outs over only a subset of training images provides the model with varied, natural-looking partial occlusions while keeping most images unchanged. Occluding just 10-30% of the sign area mimics real traffic scenes yet still leaves enough visual signal for the detector to learn the class. Applying this augmentation in roughly one-quarter of the mini-batches regularizes the network but preserves many clean examples, so accuracy on fully visible signs is not sacrificed.
The policy of overlaying an opaque mask that hides 30-60% of every sign in every image hides too much information and shifts the training distribution too far toward heavy occlusion. The approach of replacing the entire bounding box of each sign with a uniform gray patch removes all class evidence, teaching the model nothing about partially visible signs. Using a generic CutMix policy that swaps half the pixels of every image is not targeted to the object of interest and can disrupt background context rather than simulate realistic occlusion.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
What is occlusion data augmentation in machine learning?
Open an interactive chat with Bash
Why is pasting realistic object cut-outs better for handling occlusions than using an opaque mask?
Open an interactive chat with Bash
What does it mean to avoid a distribution shift during data augmentation?