multi-crop testing
**Multi-crop testing** is the **evaluation method that runs inference on several spatial crops of the same image and combines predictions to reduce framing bias** - this is especially useful when important objects are not centered or occupy only a small image region.
**What Is Multi-Crop Testing?**
- **Definition**: Inference over a predefined set of crops, often center plus four corners, followed by prediction averaging.
- **Purpose**: Ensure model sees alternative spatial contexts that one center crop may miss.
- **Common Setup**: Five-crop or ten-crop protocol depending on benchmark strictness.
- **Output Fusion**: Mean logits or probabilities across crop predictions.
**Why Multi-Crop Testing Matters**
- **Coverage**: Captures objects near edges that center crop can truncate.
- **Accuracy Gain**: Often provides incremental but reliable metric improvement.
- **Evaluation Fairness**: Reduces dependence on a single crop convention.
- **Model Diagnostics**: Reveals sensitivity to object position and framing.
- **Deployment Option**: Can be enabled for high confidence applications.
**Crop Protocols**
**Five-Crop**:
- Four corners plus center.
- Balanced cost and benefit.
**Ten-Crop**:
- Five-crop plus horizontal flips.
- Higher accuracy at higher compute cost.
**Adaptive Crop**:
- Generate crops based on saliency or detector proposals.
- Useful for objects with uncertain location.
**How It Works**
**Step 1**: Generate crop set from input image at chosen scale and run each crop through the model.
**Step 2**: Average predictions and output final class distribution, optionally with uncertainty score from crop variance.
**Tools & Platforms**
- **torchvision transforms**: Built in five-crop and ten-crop utilities.
- **timm eval scripts**: Support multi-crop validation out of the box.
- **Inference services**: Batch crops together to reduce latency overhead.
Multi-crop testing is **a simple evaluation ensemble that improves spatial robustness by checking multiple viewpoints of the same image** - it is an effective option when slight extra inference cost is acceptable.