Image Tiling for Small-Object Detection Across Image Pyramids
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learned models struggle to accurately detect both small and large objects within the same image, often missing or misidentifying small objects.
Innovation Solution
The use of an image pyramid with multiple levels at different resolutions, divided into tiles, and a machine-learned object detection model to generate bounding boxes for each tile, followed by mapping and pruning the detection data to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine-learned model processes the entire image at once, then large objects can be detected, but small objects are often missed or misidentified
Solution Approach 1:
The image is divided into multiple overlapping tiles or patches, allowing the model to process smaller regions independently. This segmentation enables the detection of small objects that would otherwise be missed in a full-image processing approach, while maintaining the ability to detect large objects through the coverage of multiple tiles.
2Measurement precision
If the image is processed at high resolution, then small objects can be detected, but computational resources increase
Solution Approach 1:
By segmenting the image into smaller tiles, the computational burden is distributed across multiple smaller processing tasks rather than one large task. This allows high-resolution processing to be applied to manageable chunks, reducing peak memory usage and enabling parallel processing that improves efficiency.
Solution Approach 2:
The method processes only relevant portions of the image at high resolution through tiling, rather than processing the entire image uniformly. Overlapping tiles provide redundant coverage that ensures small objects are captured without requiring the entire image to be processed at maximum resolution simultaneously.
3Measurement precision
If multiple tiles are processed independently, then small object detection improves, but redundant and partial bounding boxes increase
Solution Approach 1:
The system incorporates a post-processing stage that uses feedback from detected bounding boxes across multiple tiles to identify and remove duplicates. By comparing bounding box positions and overlaps, the system can distinguish between genuine multiple detections and redundant detections from overlapping tiles, eliminating false positives while preserving true detections.
4Measurement precision
If overlapping tiles are used, then small objects at tile boundaries are detected, but the number of tiles and processing time increases
Solution Approach 1:
Overlapping tiles provide excessive coverage of boundary regions, ensuring that objects spanning tile boundaries are captured in multiple tiles. This redundant coverage is intentionally accepted as a trade-off for improved detection accuracy, with the understanding that post-processing will eliminate duplicate detections.
Data Source
AI summary
A computing system for detecting objects in an image can perform operations including generating an image pyramid that includes a first level corresponding with the image at a first resolution and a second level corresponding with the image at a second resolution. The operations can include tiling the first level and the second level by dividing the first level into a first plurality of tiles and the second level into a second plurality of tiles; inputting the first plurality of tiles and the second plurality of tiles into a machine-learned object detection model; receiving, as an output of the machine-learned object detection model, object detection data that includes bounding boxes respectively defined with respect to individual ones of the first plurality of tiles and the second plurality of tiles; and generating image object detection output by mapping the object detection data onto an image space of the image.