Dataset splitting divides a labeled dataset into separate subsets used for different stages of model development. The standard split is three-way: a training set (typically 70-80% of data) that the model learns from, a validation set (10-15%) used to tune hyperparameters and monitor overfitting during training, and a test set (10-15%) held out until final evaluation to give an unbiased performance estimate.

How you split matters as much as the ratios. Random splitting works for large, diverse datasets but can leak information in datasets with similar images (consecutive video frames, multiple crops from one scene). Stratified splitting preserves the class distribution in each subset, which is important when some classes are rare. For time-series or video data, temporal splits (train on earlier data, test on later) prevent future data from leaking into training. K-fold cross-validation rotates through multiple train/val splits to get more robust performance estimates, though it's computationally expensive for large vision datasets.

A common mistake is tuning the model on the test set (directly or indirectly) by evaluating it repeatedly and making decisions based on test results. This inflates reported performance and leads to surprises in production. The test set should be evaluated once, at the end.