Pseudo Label Selection for Small-Object Detection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods using pseudo labels for unlabeled data often discard pseudo labels of small object images, leading to an imbalance in image sizes and decreased detection accuracy for small objects.
Innovation Solution
A learning method that selects pseudo labels for use in machine learning based on both reliability and object image size, ensuring balanced representation of small and large objects during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pseudo labels are selected based solely on reliability, then high-reliability pseudo labels are prioritized, but pseudo labels of small object images are discarded leading to detection accuracy degradation for small objects
Solution Approach 1:
The patent applies local quality by differentiating the selection criteria for pseudo labels based on object size. Small object pseudo labels are given preferential treatment in the selection process, allowing them to be retained even with lower confidence scores, while large object pseudo labels follow the traditional confidence-based filtering. This localized adjustment ensures that small objects, which are more critical and harder to detect, receive adequate representation in the training data without compromising overall reliability
Solution Approach 2:
The patent introduces asymmetry in the pseudo label selection process by treating small and large objects differently. Instead of applying a uniform confidence threshold to all objects, the system applies asymmetric selection rules: small objects are retained with lower thresholds and prioritized in the selection queue, while large objects follow standard reliability-based filtering. This asymmetric approach corrects the imbalance that would otherwise occur in the training data distribution
2Reliability
If pseudo labels of small object images are discarded to maintain high reliability, then overall model reliability is improved, but imbalance in object size distribution occurs in learning data
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the confidence threshold parameter based on object size. For small objects, a lower confidence threshold is applied, allowing more small object pseudo labels to be retained in the training set. For large objects, a higher confidence threshold is maintained. This parameter adaptation ensures that the quantity of small object samples in the training data is sufficient for learning, while still maintaining acceptable reliability standards
Solution Approach 2:
The patent implements preliminary action by pre-classifying pseudo labels according to object size before the selection process. Small object pseudo labels are identified and placed in a prioritized queue before confidence-based filtering is applied. This preliminary classification ensures that small objects are considered first in the selection process, preventing their premature discarding and ensuring adequate representation in the final training set
3Reliability
If confidence-based filtering is applied to all pseudo labels uniformly, then high-confidence pseudo labels are selected, but small object pseudo labels are disproportionately discarded
Solution Approach 1:
The patent applies local quality by implementing size-adaptive confidence thresholds. Instead of using a single uniform threshold for all objects, the system adjusts the threshold locally based on object size characteristics. Small objects, which inherently have lower detection confidence due to their size, are evaluated against a lower threshold, while large objects use a higher threshold. This local adaptation ensures that the confidence filtering process is equitable across different object sizes
Solution Approach 2:
The patent introduces dynamics by making the selection process adaptive rather than static. The confidence threshold and selection probability are dynamically adjusted based on the detected object size. Small objects receive dynamic preferential treatment with lower thresholds and higher retention probabilities, while large objects follow the standard dynamic filtering process. This dynamic approach allows the system to adapt to the specific characteristics of each object class and size category
Data Source
AI summary
A learning method for performing learning of a machine learning model using unlabeled data with no labels includes: inputting the unlabeled data to the machine learning model to generate pseudo labels; performing a first selection of selecting a pseudo label for learning from the generated pseudo labels based on reliability; performing a second selection of selecting, based on image sizes of objects to which the pseudo labels are given, the pseudo label for learning from pseudo labels that are discard targets that has not been selected as the pseudo label for learning in the first selection; and performing the learning using the pseudo label for learning.