Neural Network Object Detection Beyond Trained Size Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Object detection neural networks are limited in detecting objects outside a certain size range without significantly increasing complexity, leading to ignored detections of smaller objects.
Innovation Solution
The method involves using prior knowledge to determine whether to approve or ignore neural network outputs based on sensed information, allowing for detection of objects outside the trained size range without increasing the neural network's complexity, by utilizing additional features generated during previous processing steps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the object detection neural network is trained to detect objects of a certain size range, then the size and complexity of the neural network is reduced, but objects outside the certain size range cannot be detected
Solution Approach 1:
The system performs preliminary actions by generating additional features during previous processing steps before final detection. These pre-computed features (such as contextual information, semantic labels, or multi-scale representations) are stored and later used to enhance detection of objects outside the trained size range, allowing the network to handle diverse object sizes without increasing its structural complexity
Solution Approach 2:
Additional features generated during previous processing steps serve as intermediaries between the neural network output and the final detection decision. These intermediate features bridge the gap between objects of different sizes, enabling the network to detect objects outside its trained size range by leveraging these intermediary representations
2Adaptability or versatility
If the neural network complexity is increased to detect objects outside the certain size range, then detection versatility is improved, but computational and storage resources increase
Solution Approach 1:
By pre-generating and storing additional features during earlier processing steps, the system avoids the need to increase neural network complexity during inference. These pre-computed features are reused for detecting objects of various sizes, reducing computational resources during actual detection while maintaining versatility
Solution Approach 2:
The system creates copies of useful features (such as multi-scale versions, contextual embeddings, or semantic representations) that can be reused across different detection scenarios. Instead of training a larger network, the system replicates and reuses these feature copies to detect objects outside the original size range efficiently
3Adaptability or versatility
If the neural network complexity is increased to detect objects outside the certain size range, then detection versatility is improved, but storage resources increase
Solution Approach 1:
The system extracts only the essential additional features needed for detecting objects outside the size range, rather than storing complete models or large datasets. By extracting and storing only the necessary feature representations (such as contextual vectors or scale-invariant descriptors), the system maintains detection versatility while minimizing storage requirements
Solution Approach 2:
Instead of storing multiple complex models for different object sizes, the system creates compact copies of key features that can be applied across size ranges. These feature copies are much smaller in storage requirement compared to full network models, yet enable detection of objects outside the trained size range
Data Source
AI summary
A method for overcoming a detection limitation of a neural network, the method includes obtaining a sensed information unit that captures an object; obtaining an indication for a detection limitation of the neural network with respect to the object, wherein the detection limitation of the neural network prevents the neural network from generating a neural network output that is indicative of the object with at least a desirable certainty; feeding the sensed information unit to the neural network to provide a neural network output; and controlling a detection of the object by the neural network based on an indication that the object is captured in the sensed information unit.