While tuning a CNN that classifies photographs of industrial defects, you observe that validation accuracy drops sharply whenever the defect is partly hidden by a worker's hand or tool, even though training loss remains low. You decide to add a masking-based data-augmentation step that follows the Random Erasing technique. Which configuration is MOST likely to increase robustness to partial occlusion without altering the ground-truth class labels?
Replace all background pixels that fall outside each annotated bounding box with a uniform gray mask so that only the foreground object remains visible.
With a fixed probability, overwrite one randomly located rectangular region covering roughly 2-20 % of every training image with random pixel values (or the per-channel dataset mean) while keeping the original label.
At each forward pass, randomly zero-out a comparable fraction of convolutional filters in the network's first layer to simulate information loss.
Add a binary channel that records the location of an arbitrary mask and train the model to reconstruct the hidden pixels as a secondary objective.
Random Erasing augments training data by selecting a random rectangle (typically 2-20 % of the image area with a random aspect ratio) and replacing its pixels with random values or the dataset's mean. The label remains unchanged, forcing the network to rely on contextual cues distributed across the image, which empirically improves robustness to occlusion and reduces over-fitting.
Replacing the entire background (second choice) changes the statistical structure of every image rather than introducing random occlusions and can even remove features that are useful for differentiation. Randomly dropping convolutional filters (third choice) is a form of network regularization (analogous to dropout) rather than data augmentation and does not teach the model to handle occluded inputs. Adding an explicit binary mask channel and reconstruction task (fourth choice) converts the problem into multitask learning and signals the network where the occlusion is, defeating the purpose of forcing invariance to hidden regions.
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 the Random Erasing technique in data augmentation?
Open an interactive chat with Bash
Why does Random Erasing improve robustness to occlusion over methods like masking the background?
Open an interactive chat with Bash
Why is multitask learning not suitable for this specific problem?