Rotation Prediction is an early self-supervised pretext task where the model is trained to predict which rotation (0°, 90°, 180°, 270°) was applied to an input image — requiring the network to learn meaningful visual features (object orientation, shape, semantics) to solve the task.

How Does Rotation Prediction Work?

• Process: Randomly rotate each image by 0°, 90°, 180°, or 270°. The network must classify which rotation was applied.
• Labels: Free (generated by the augmentation, no human annotation needed).
• Architecture: Standard CNN (e.g., ResNet) + 4-class classification head.
• Paper: RotNet (Gidaris et al., 2018).

Why It Matters

• Simplicity: One of the simplest and most effective early pretext tasks.
• Insight: To predict rotation, the network must understand "up" vs. "down" and object semantics — non-trivial!
• Legacy: Largely superseded by contrastive methods (SimCLR, MoCo, DINO) but remains a pedagogical benchmark.

Rotation Prediction is the compass test for neural networks — a deceptively simple pretext task that requires genuine visual understanding to solve.