Delivering image tiles and YOLO-format bounding box annotations specifically processed for small object detection. It features sliced image segments designed to preserve spatial resolution for targets that occupy a small percentage of the original frame, preventing feature loss during standard neural network downsampling.
Use Cases
- Train a YOLO-based detector using the class_id and normalized bounding box coordinates in the label files
- Evaluate the impact of image slicing on mean Average Precision (mAP) for small object classes
- Develop post-processing scripts to merge tile-level detections back into full-frame coordinates
Strengths
- Includes image tiles (slices) specifically generated for small object detection tasks
- Provides annotations in the standard YOLO text format (.txt)
- Focuses on increasing the pixel-to-object ratio for tiny targets within high-resolution imagery