Automated Prior Selection for Detection CNNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detector architectures, such as SSD, require manual configuration of priors, which is knowledge-intensive and does not consider detection quality metrics, leading to suboptimal performance and increased computational load.
Innovation Solution
An automated system for selecting priors, which generates a set of redundant priors, trains a CNN, and selects a subset based on detection quality metrics, reducing the number of priors and feature maps while maintaining accuracy, thereby decreasing computational load and improving detection speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual configuration of priors is used, then expert knowledge can be applied, but the process is knowledge-intensive and does not consider detection quality metrics
Solution Approach 1:
The system automatically selects priors based on detection quality metrics without requiring manual expert configuration. The automated selection process evaluates multiple candidate priors and chooses the optimal subset based on actual detection performance, eliminating the need for knowledge-intensive manual tuning while improving detection quality.
Solution Approach 2:
The system uses detection quality metrics as feedback to iteratively improve prior selection. By evaluating detection quality with candidate priors and using this feedback to select the optimal subset, the system continuously optimizes performance based on actual detection outcomes rather than relying on manual expert judgment.
2Reliability
If a large number of priors are used, then detection coverage is improved, but computational load increases
Solution Approach 1:
The system extracts and selects only the most useful subset of priors from a larger pool of candidate priors. By identifying and removing redundant or less effective priors based on detection quality metrics, the system maintains detection accuracy while significantly reducing the number of priors that need to be processed, thereby lowering computational load.
Solution Approach 2:
The system changes the parameter of prior quantity from a fixed large number to an optimized subset size. By dynamically adjusting the number of priors based on detection quality evaluation, the system finds the optimal balance between detection accuracy and computational efficiency, using fewer priors than traditional approaches while maintaining or improving performance.
3Measurement precision
If CNN retraining is performed in each pruning iteration, then model optimization is achieved, but the process becomes extremely time-consuming
Solution Approach 1:
The system performs preliminary evaluation of candidate priors using detection quality metrics before final model training. By assessing prior effectiveness in advance using the pre-trained CNN and selecting the optimal subset beforehand, the system avoids repeated full CNN retraining iterations, significantly reducing the time required for model optimization while still achieving detection accuracy improvements.
Data Source
AI summary
An example apparatus for selecting priors includes a training set receiver to receive a training dataset. The apparatus includes a prior generator to generate a set of redundant priors based on the training dataset. The apparatus includes an intermediate trainer to train a detection CNN using the set of redundant priors. The apparatus includes a score and location receiver to send all training samples of the training dataset to the trained detection CNN and receive responses for all of the redundant priors in the set of redundant priors. The apparatus includes a subset selector to select a subset of the set of redundant priors based on the responses.