An ML engineer trained an accident-detection model using Data Wrangler and achieved strong training/validation results, but production performance suffers due to varied camera image quality. Which change will improve accuracy fastest?

The correct answer is to recreate the training dataset in Data Wrangler using the corrupt image transform with the impulse-noise option. This directly addresses the problem of varied camera image quality by introducing noise into the training data, making the model more robust to real-world imperfections. This data augmentation technique helps the model learn to identify accidents even with noisy or degraded input, improving generalization. Collecting additional images from every camera is a valid long-term strategy but is time-consuming and may not be the fastest way to improve accuracy, especially if the new images still suffer from quality variations. Enhancing image contrast might help with some quality issues but doesn't specifically target noise or general degradation. Resizing images to a uniform size is often a preprocessing step but doesn't directly address the model's robustness to varied image quality or noise.