Online Tracking Inference Scheduling After Frame Training Interruptions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object tracking systems using machine learning face interruptions during frame periods, leading to loss of tracking state due to incomplete training, especially when processing is interrupted by user operations or outlier inputs, and require restarting training from an initial state.
Innovation Solution
An information processing apparatus and method that executes inference and training processes in subsequent frame periods to ensure complete training before inference, using a convolutional neural network with online training to update coefficients continuously, even after interruptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If online training is performed every frame to maintain tracking accuracy, then tracking precision is improved, but processing time increases and interruptions occur
Solution Approach 1:
The patent applies preliminary action by completing the training process for frame N-1 before executing the inference process for frame N. This ensures that training is always up-to-date before inference, maintaining tracking accuracy without requiring simultaneous execution of both processes during the same frame period, thus avoiding processing delays and interruptions.
2Ease of operation
If processing is interrupted by user operations or outlier inputs, then system responsiveness is improved, but tracking state is lost
Solution Approach 1:
The patent implements beforehand cushioning by maintaining a buffer mechanism where the training result from frame N-1 is preserved and can be restored after interruptions. This cushioning effect ensures that when processing is interrupted by user operations or outlier inputs, the tracking state is not lost but can be recovered, maintaining reliability while allowing system responsiveness.
3Reliability
If training is completed before inference for each frame, then tracking reliability is improved, but processing complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the processing into distinct sequential stages: training for frame N-1 is completed in one segment, then inference for frame N is executed in the next segment. This segmentation simplifies the overall processing logic by avoiding the complexity of concurrent training and inference operations, while still achieving the goal of having up-to-date training results available for reliable tracking.
Data Source
AI summary
An information processing apparatus comprises a processing unit configured to execute, in frame periods corresponding to respective frames, an inference relating to the frame and training relating to the frame. When an inference relating to a first frame has been completed but training relating to the first frame has not been completed in a first frame period corresponding to the first frame, the processing unit executes training relating to the first frame and an inference and training relating to a second frame in a second frame period corresponding to the second frame subsequent to the first frame.