Object Detection With Bounding-Box Pixel Contributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models, particularly deep-learning-based models, lack the ability to clearly explain the basis for object detection, making it difficult for users to understand the contributions of pixels in an input image, and the gradient of bounding boxes cannot be calculated due to NMS not being represented by a partial differential equation.
Innovation Solution
An object detecting apparatus that includes a box generating unit to detect objects and generate bounding boxes, a box selecting unit to select the most reliable bounding box, an overlap computing unit to calculate overlaps, and a contribution computing unit to determine pixel contributions based on these overlaps, using techniques like Grad-CAM to visualize pixel contributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep-learning-based image recognition models are used for object detection, then detection accuracy is improved, but understandability of detection bases deteriorates
Solution Approach 1:
The patent introduces pixel contribution information as an intermediary between the deep-learning model and the user. This intermediary element (pixel contribution data) bridges the gap by providing explanatory information about which pixels influenced the detection result, making the black-box model's decision process transparent without altering the model's high accuracy performance.
2Reliability
If NMS is applied to select reliable bounding boxes, then detection reliability is improved, but calculability of pixel contributions deteriorates
Solution Approach 1:
The patent performs preliminary action by calculating pixel contribution information for all candidate bounding boxes before NMS is applied. By pre-calculating these contributions and storing them, the system avoids the complexity of calculating partial derivatives of NMS later, enabling straightforward retrieval and use of pixel contribution data for the finally selected bounding box.
3Loss of information
If gradient calculation is performed for pixel contribution analysis, then explainability is improved, but computational complexity deteriorates
Solution Approach 1:
The patent performs the computationally intensive gradient calculation and pixel contribution analysis in advance, before the final object detection result is needed. This preliminary computation stores the explainability data in a readily accessible format, avoiding the need for repeated complex calculations when users request explanation of detection results.
Data Source
AI summary
An object detecting apparatus includes processing circuitry configured to; detect an object in an input image and generate one or more bounding boxes enclosing the object; select, if a plurality of mutually overlapping bounding boxes is generated, one bounding box from the plurality of mutually overlapping bounding boxes on a basis of a reliability of each of bounding boxes; calculate an overlap between each of bounding boxes having been generated and the bounding box having been selected; and calculate a contribution of a pixel contributing to the detection of the object in a plurality of pixels included in the input image on a basis of the calculated overlap.