A data science team is developing a computer vision model to classify traffic signs, including those with directional arrows (e.g., 'turn left,' 'keep right'). To expand the training dataset and improve model robustness, the team plans to use flipping as a data augmentation method. Which application of flipping poses the most significant risk of introducing label-image mismatches that could degrade the model's real-world accuracy?
Applying vertical flipping only to asymmetrical signs (e.g., signs with arrows).
Applying vertical flipping across the entire dataset.
Applying horizontal flipping only to symmetrical signs (e.g., 'Stop' signs).
Applying horizontal flipping across the entire dataset.
The correct answer is applying horizontal flipping across the entire dataset. Data augmentation techniques like flipping must be applied with careful consideration of the data's context. For images where left-right orientation is meaningful, such as a traffic sign with a 'turn left' arrow, a horizontal flip will create an image that directs a 'turn right'. If the original label ('turn left') is not also changed, the model will be trained on corrupted data, learning that a right-turning arrow means 'turn left', which will degrade its performance on real-world data.
Applying vertical flipping is less risky. While traffic signs are rarely seen upside down, a vertically flipped sign is usually just an anomaly and does not transform into a different, valid instruction.
Applying horizontal flipping only to symmetrical signs is a valid and safe augmentation strategy. It correctly avoids flipping signs where direction matters, and thus is a risk mitigation technique, not a risk itself.
Applying vertical flipping only to asymmetrical signs carries the same low risk as applying it to all signs; it does not create label-image mismatches that represent valid but incorrect instructions.
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.
Why is horizontal flipping across the dataset risky for directional traffic signs?
Open an interactive chat with Bash
Why doesn’t vertical flipping pose a similar risk to horizontal flipping?
Open an interactive chat with Bash
What makes flipping symmetrical signs safe for data augmentation?