multi-scale testing

**Multi-Scale Testing** is a **test-time technique that runs inference at multiple input resolutions and combines the results** — detecting objects or segmenting scenes more accurately by capturing features at different spatial scales. **How Does Multi-Scale Testing Work?** - **Scales**: Resize the input to multiple resolutions (e.g., 0.5×, 0.75×, 1.0×, 1.25×, 1.5×). - **Infer**: Run the model at each scale independently. - **Combine**: Average the predictions (for segmentation) or merge detections (NMS for detection). - **Optional**: Combine with horizontal flipping for additional views. **Why It Matters** - **Object Size Variation**: Small objects are better detected at larger scales. Large objects at original scale. - **Segmentation**: Multi-scale testing consistently improves mIoU by 1-3% on semantic segmentation benchmarks. - **Competitions**: Standard practice in segmentation and detection competitions (but too slow for real-time). **Multi-Scale Testing** is **seeing at every zoom level** — running inference at multiple resolutions to capture objects and details at all spatial scales.