Object Detection Training Sample Selection for Small Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models exhibit low detection accuracy for small objects due to characteristics such as motion blur and susceptibility to occlusion, leading to inconsistent sample selection and inadequate feature learning.
Innovation Solution
Perform movement operations on object regions in sample images to enhance object regions, determining a set of training samples based on candidate regions to improve the selection of small object samples without significantly affecting large object samples, thereby training an object detection model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If movement operations are performed on all object regions to enhance training samples, then the detection accuracy of small objects is improved, but the processing complexity and time consumption increase
Solution Approach 1:
The patent applies different processing strategies to different object regions based on their size characteristics. Small object regions undergo movement operations (translation, rotation, scaling) to generate diverse training samples, while large object regions use traditional selection methods. This localized differentiation improves small object detection without unnecessarily complicating the processing of large objects.
Solution Approach 2:
The patent segments the object region selection process into distinct stages: initial candidate region generation, small object region identification, movement operation application, and final training sample determination. This segmentation allows complex processing to be applied only where necessary (small objects) while maintaining efficiency for other cases.
2Reliability
If movement operations are applied to generate more training samples, then the representation of small objects in training data is improved, but the training time and computational resources increase
Solution Approach 1:
The patent performs movement operations during the training sample preparation phase before actual model training begins. By pre-generating enhanced training samples with varied positions, orientations, and scales through translation, rotation, and scaling operations, the model learns more robust small object features without incurring additional computational overhead during the training execution phase.
3Measurement precision
If the selection threshold for training samples is lowered to include more small objects, then the detection accuracy improves, but the inclusion of negative samples increases
Solution Approach 1:
The patent introduces an intermediary verification mechanism using overlap degree calculation between candidate regions and ground truth labels. Before including a candidate region as a positive training sample, the system calculates the overlap degree to verify genuine object presence. This intermediary step filters out negative samples that would otherwise be incorrectly included due to low selection thresholds, maintaining high precision while enabling aggressive sampling.
Data Source
AI summary
A method and an apparatus for determining a set of training samples, a method and an apparatus for training a model, and a method and an apparatus for detecting an object, the method including: performing a movement operation on an object region in a sample image to determine a plurality of enhanced object regions; and determining, based on a plurality of candidate regions corresponding to the sample image and the plurality of enhanced object regions, a set of training samples corresponding to the sample image to obtain an object detection model by training a network model based on the set of training samples corresponding to the sample image. A probability of a small object region being selected as a set of training samples is improved without affecting sample selection of a large-size object region, thereby improving detection accuracy of an object detection model in detecting the small object.