Instance Segmentation Training for Proximate Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning (ML) models for computer vision struggle with accurately distinguishing between proximate objects and require excessive computation, making them unsuitable for real-time applications like autonomous vehicle navigation.
Innovation Solution
A method involving training ML models with improved hard example mining and non-maximum suppression techniques, along with multi-stage training using focal loss functions and constrained receptive fields, to enhance ROI accuracy and reduce computation time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ML models use standard training methods to identify objects, then object detection capability is achieved, but the models cannot properly distinguish between proximate objects, resulting in inaccurate identifications
Solution Approach 1:
The patent segments the object identification process into multiple stages: generating multiple candidate regions of interest (ROIs) for each detected object, evaluating each ROI independently through multiple ML models, and selecting the best ROI based on confidence scores. This segmentation allows the system to distinguish between proximate objects by treating their potential overlapping regions separately rather than as a single undifferentiated detection.
Solution Approach 2:
The patent generates multiple candidate ROIs for each detected object, including some that may partially overlap or extend beyond the actual object boundaries. This excessive generation of candidate regions ensures that the true object region is captured among the candidates, even if it requires additional evaluation steps to identify the correct one. The system evaluates multiple hypotheses rather than relying on a single detection.
2Measurement precision
If ML models provide accurate object identifications, then detection precision is improved, but compute time increases to over 100 milliseconds, making them unsuitable for real-time applications
Solution Approach 1:
The patent performs preliminary filtering of candidate ROIs using confidence score thresholds before subjecting them to full ML model evaluation. ROIs with low confidence scores are discarded early in the process, preventing unnecessary computation on unlikely candidates. This preliminary action reduces the number of expensive ML model evaluations required while maintaining detection accuracy for high-confidence objects.
Solution Approach 2:
The patent applies different evaluation strategies to different candidate ROIs based on their local characteristics. High-confidence ROIs undergo simpler validation, while low-confidence or ambiguous ROIs receive more rigorous multi-model evaluation. This local differentiation of processing quality allows the system to allocate computational resources efficiently, spending more time on difficult cases and less on straightforward detections.
3Measurement precision
If ML models process images to identify objects, then object detection is achieved, but the detected object may have moved significantly by detection time, rendering the detection unreliable for decision-making
Solution Approach 1:
The patent implements periodic processing of image frames at high frame rates (e.g., 20-30 fps or higher), creating multiple detection opportunities as objects move through the scene. By processing images periodically rather than waiting for a single comprehensive analysis, the system can detect objects at multiple time points, increasing the likelihood of obtaining an accurate, up-to-date detection that reflects the object's current position for reliable decision-making.
Data Source
Figure 1
Figure 2A~2B
Figure 3
AI summary
Improved techniques for training a machine learning (ML) model are discussed herein. Training the ML model can be based on a subset of examples. In particular, the training can include identifying a reference region associated with an area of the image representing an object, and selecting, based at least in part on a first confidence score associated with a first bounding box, a first hard example for inclusion in the subset of examples. In some cases, the first confidence score and the first bounding box can be associated with a first portion of the feature map. Next, the training can include determining that a first degree of alignment of the first bounding box to the reference region is above a threshold degree of alignment, and in response, replacing the first hard example with a second hard example.