Object Detection Model Retraining via Redundant Frame Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection systems face challenges in achieving high accuracy due to poor training data quality, leading to false positives and negatives, particularly in surveillance and retail analytics applications, where redundant data can increase the likelihood of detection errors.
Innovation Solution
A method is introduced that involves identifying frames where the ROI detection model fails to detect objects or generates false positives, using a hybrid approach combining ROI detection, tracking, and motion modeling to select and re-train the model with diverse data sets, including frames with occlusions, varying lighting, and different camera perspectives, to improve model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If all generated training data from on-premise raw videos is used, then the training dataset size increases, but the model accuracy decreases due to redundant data causing false positives and false negatives
Solution Approach 1:
The patent extracts and removes redundant data from the training dataset by comparing consecutive video frames and identifying duplicate or highly similar frames. This extraction process eliminates redundant information that would otherwise cause the model to learn incorrect patterns and produce false positives/negatives, while preserving the essential diverse training samples needed for accurate detection.
Solution Approach 2:
The patent changes the parameter of data selection by introducing frame similarity thresholds and redundancy metrics. Instead of using all generated training data uniformly, the system applies parameter-based filtering to selectively include only high-quality, non-redundant frames in the training dataset, thereby improving model accuracy while maintaining an adequate dataset size.
2Adaptability or versatility
If the ROI detection model is trained with diverse data including occlusions and varying lighting, then the model adaptability improves, but the training complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-processing video frames to detect and annotate diverse scenarios (occlusions, varying lighting conditions, different camera perspectives) before training. By identifying and preparing these challenging cases in advance, the system incorporates them systematically into the training dataset without overwhelming complexity during the actual model training phase.
Solution Approach 2:
The patent creates a universal training framework that handles multiple diverse scenarios (occlusions, lighting variations, different perspectives) through a single integrated data collection and processing pipeline. This multi-functional approach allows the model to learn from various conditions simultaneously, improving adaptability while managing training complexity through unified processing.
Data Source
AI summary
Example implementations include a method, apparatus, and computer-readable medium for object detection, comprising detecting a first object in a first image frame and a second image frame, wherein the first object is bounded by region-of-interest (ROI) boundaries generated by an ROI detection model. The implementations further include calculating a speed of the first object using positions of the first object in the first and second image frame, identifying at least one image frame that should include the first object based on a calculated speed of the first object. The implementations further include determining that the at least one image frame should be added to a training dataset for the ROI detection model in response to detecting that the ROI detection model did not generate a ROI boundary in the at least one image frame, and subsequently re-training the ROI detection model using said training dataset.