Multi-Scale Object Inference for Real-Time Detection Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning models for computer vision in applications like autonomous vehicles face challenges in accurately detecting objects in real-time due to high computational requirements and flawed object identification, particularly when objects are proximate or of varying sizes, leading to unreliable decision-making.
Innovation Solution
The techniques involve training machine-learning models using improved hard example mining, non-maximum suppression replacement, and multi-scale inference to enhance the accuracy and speed of object detection by selecting specific examples for training, determining ROI sizes for optimal performance, and using focal loss functions to weight errors effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-accuracy ML models are used for object detection, then detection precision improves, but computational time increases beyond real-time requirements
Solution Approach 1:
The patent divides the object detection task into multiple stages with different levels of detail. A coarse detection stage identifies potential objects quickly, followed by a refined detection stage that applies more computationally intensive processing only to regions of interest. This segmentation allows the system to achieve high accuracy for critical detections while maintaining real-time performance by avoiding exhaustive processing of the entire image.
Solution Approach 2:
The patent applies full computational resources selectively rather than uniformly across all detections. High-accuracy processing is applied only to objects that meet certain criteria (e.g., confidence thresholds, object importance), while other objects receive streamlined processing. This partial application of intensive computation maintains real-time performance while achieving high accuracy where needed.
2Reliability
If comprehensive ML model processing is applied to all image portions, then detection coverage improves, but processing speed decreases for real-time applications
Solution Approach 1:
The patent applies different processing qualities to different regions of the image based on their importance. Regions containing critical objects or areas with higher uncertainty receive more intensive processing and higher confidence thresholds, while less critical regions use streamlined processing. This local differentiation maintains comprehensive coverage for safety-critical detections while accelerating processing for less important areas.
Solution Approach 2:
The system performs comprehensive processing only when necessary rather than uniformly for all image portions. By identifying regions of interest through initial screening and applying detailed processing only to those regions, the system achieves reliable detection coverage for critical objects while maintaining high processing speeds overall.
3Adaptability or versatility
If ML models process objects of all sizes uniformly, then detection versatility improves, but accuracy decreases for proximate or varying-sized objects
Solution Approach 1:
The patent adapts processing parameters locally based on object characteristics such as size, position, and type. Different detection thresholds, confidence levels, and processing intensities are applied to different object classes and size ranges. For example, proximate objects receive adjusted processing to resolve spatial relationships, while varying-sized objects receive scale-appropriate processing parameters, thereby maintaining high accuracy across diverse detection scenarios.
Data Source
AI summary
Techniques for utilizing multiple scales of images as input to machine learning (ML) models are discussed herein. Operations can include providing an image associated with a first scale to a first ML model. An output of the first ML model can include a first bounding box indicative of a first region of the image representing a first object, with the first bounding box falling within a first range of sizes. Next, a scaled image can be generated by scaling the image. The scaled image can be provided to a second ML model, which can output a second bounding box indicative of a second region of the image representing a second object, the second bounding falling within a second range of sizes. Thus, inputting a scaled image to a same ML model (or to different ML models) can result in different detected features in the images.