Object Detection Size Feature Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods in computer vision struggle with accurate size prediction of objects in images due to features extracted by deep learning networks being of identical length, regardless of object size, leading to inaccurate size estimation.
Innovation Solution
The proposed method involves generating size-related features by comparing the enclosing frame of a target object with a predetermined candidate frame set of varying sizes, selecting reference frames based on overlap degrees, and combining these size-related features with original features in a machine learning model for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning network extracts features from candidate areas, then feature extraction is automated and detection framework is end-to-end, but all extracted features have identical length which lacks correlation with object size leading to inaccurate size prediction
Solution Approach 1:
The patent segments the feature extraction process into multiple components: original features from deep learning network and size-related features from candidate frame set comparison. This segmentation allows each component to serve its specific function while together they provide comprehensive information for accurate detection.
Solution Approach 2:
The patent adds a new dimension to feature representation by incorporating size-related features derived from comparing enclosing frames with candidate frames of varying sizes. This transforms the feature space from fixed-length vectors to enriched feature sets that include size correlation information.
2Adaptability or versatility
If features are designed manually with shallow classifier framework, then feature design flexibility is high, but detection framework is not end-to-end and requires more manual intervention
Solution Approach 1:
The patent merges manual feature design capabilities with automated deep learning feature extraction. By combining original features from the neural network with size-related features from frame comparison, it achieves both automation and design flexibility in a unified end-to-end framework.
3Productivity
If only original features from deep learning network are used, then feature extraction is simple and fast, but size prediction accuracy is insufficient due to lack of size correlation
Solution Approach 1:
The patent performs preliminary action by pre-defining a candidate frame set with various sizes before the actual detection process. This allows size-related features to be quickly computed during detection by comparing enclosing frames against the pre-prepared candidate set, maintaining speed while improving accuracy.
Data Source
Figure 1~2
Figure 3A~3B
Figure 3C~5
AI summary
An object detection method, device, apparatus and a computer-readable storage medium are provided. The method includes: obtaining an enclosing frame of a target object in an input image; according to the enclosing frame, determining a reference frame from a predetermined candidate frame set comprising a plurality of candidate frames; generating a size-related feature according to a size of the reference frame and a size of the enclosing frame; and detecting an object in the input image by applying the size-related feature in a machine learning model. In an embodiment of the present application, the object detection is performed by using a feature related to an object size, that is, the prediction criterion related to the object size is added to an original feature in a machine learning model, thereby further improving the accuracy of the object detection.