Object Detection Model Retraining via Classification Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection systems face challenges in achieving high accuracy due to poor performance on pre-trained models when training data is inadequate, leading to false positives and negatives, and are unable to effectively select and utilize relevant data for re-training.
Innovation Solution
A method involving a classification tracking model that identifies frames where objects are missed or incorrectly detected, adding these frames to a training dataset and re-training the model using a hybrid approach combining ROI detection, tracking, and motion modeling to improve data selection and model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If all generated training data is used for model training, then the training dataset size increases, but the model accuracy deteriorates due to redundant data causing false positives and false negatives
Solution Approach 1:
The system extracts and removes redundant data from the training dataset using a determination model that identifies duplicate or similar frames. By taking out only the necessary unique frames and eliminating redundant ones, the system maintains model accuracy while using a smaller, more effective training dataset.
Solution Approach 2:
The system applies different quality standards to different portions of the training data. Instead of treating all generated data uniformly, it evaluates each frame locally using the determination model to assess its uniqueness and value, then selectively includes only high-quality, non-redundant frames in the retraining process.
2Device complexity
If pre-trained models are used without re-training, then the system complexity is reduced, but the detection accuracy deteriorates due to poor performance on specific on-premise scenes
Solution Approach 1:
The system performs preliminary evaluation of the pre-trained model's performance on specific on-premise scenes before committing to full retraining. By using the determination model to assess whether retraining is actually needed based on scene characteristics and current model performance, it avoids unnecessary retraining operations while maintaining high detection accuracy when required.
Solution Approach 2:
The system dynamically adjusts between using pre-trained models and re-training based on runtime conditions. The determination model continuously evaluates whether the current pre-trained model is sufficient for the specific on-premise scene, allowing the system to switch between the two approaches optimally rather than being static.
3Measurement precision
If manual data selection is performed for re-training, then the model accuracy improves through curated training data, but the operation time and labor increase significantly
Solution Approach 1:
The system implements self-service automated data selection using the determination model that automatically evaluates generated training data, identifies redundant frames, and selects optimal frames for retraining without human intervention. This automated process maintains high model accuracy while eliminating the time-consuming manual data curation process.
Solution Approach 2:
The system uses feedback from the determination model's analysis of training data quality to automatically adjust the data selection process. By continuously monitoring which frames contribute most to model performance and which are redundant, the system refines its automated selection criteria, improving accuracy while minimizing manual oversight requirements.
Data Source
AI summary
Disclosed herein is an object detection system, including apparatuses and methods for object detection. An implementation may include receiving a first class of a first object depicted in an image frame from a classification model and subsequently receiving a second image frame. The implementation further includes predicting, using a classification tracking model, that the classification model will output the first class for the second image frame and then detecting whether the first class is in fact outputted. The implementation includes determining that the second image frame should be added to a training dataset for the classification model when detecting that the classification model did not generate the first class for the second image frame as predicted and re-training the classification model using the training dataset.