Multimodal Position Learning Using Reference-Object Spatial Loss

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques struggle to accurately infer a target position using an image and text that designates the target position by referring to an object within the image.

Innovation Solution

A learning apparatus that acquires teaching data comprising input images and texts, generates output data to specify the target position and its relationship to a reference object, and updates model parameters using a loss function based on errors between the output and correct answer data to enhance accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a model is trained using conventional loss functions that only consider target position error, then the training process is simple, but the target position inference accuracy is insufficient

Engineering Contradiction:
Improvetarget position inference accuracyVSAvoidloss function complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The loss function is segmented into three distinct error components: target position error (first error), reference position error (second error), and positional relationship error (third error). Each component independently evaluates a specific aspect of the prediction, allowing the model to learn multiple spatial relationships simultaneously. This segmentation resolves the contradiction by breaking down the complex accuracy requirement into manageable error components that can be optimized separately yet contribute to overall inference accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The invention extends the traditional single-dimension target position error metric to a multi-dimensional error evaluation space. By incorporating reference position and positional relationship errors as additional dimensions, the loss function comprehensively evaluates spatial predictions from multiple perspectives. This dimensional expansion resolves the contradiction by providing a more complete accuracy metric without overwhelming complexity, as each dimension addresses a specific aspect of spatial understanding.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Reliability

If a loss function based on multiple errors is used to improve target position inference accuracy, then the model performance improves, but the computational complexity increases

Engineering Contradiction:
Improvemodel accuracyVSAvoidcomputational resources
Core Design Contradiction:
ReliabilityVSPower

Solution Approach 1:

The loss function implements partial action by selectively focusing on three critical error components (target position, reference position, and positional relationship) rather than attempting to minimize all possible errors. This selective approach achieves reliable target position inference by concentrating computational resources on the most impactful error sources, resolving the contradiction between accuracy improvement and computational complexity through targeted optimization.

Inventive Principle:
Principle #16Partial or excessive action

3Adaptability or versatility

If only target position error is minimized during training, then the training process is efficient, but the model cannot accurately understand spatial relationships

Engineering Contradiction:
Improvespatial relationship understandingVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The loss function merges three previously separate evaluation criteria (target position accuracy, reference position accuracy, and positional relationship accuracy) into a unified multi-component loss function. This merging allows the model to learn spatial relationships comprehensively during a single training process, resolving the contradiction by integrating multiple learning objectives into one cohesive training framework that improves adaptability without requiring separate training runs.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4730218A1Learning device, learning method, and program
Publication Date: 2026.04.22 HONDA MOTOR CO LTD
  • EP4730218A1 patent drawingFigure 1
  • EP4730218A1 patent drawingFigure 2
  • EP4730218A1 patent drawingFigure 3

AI summary

A learning apparatus for performing machine learning includes an acquisition unit for acquiring teaching data including input data and correct answer data, the input data including an input image that contains a reference object and an input text that relatively designates a target position by referring to the reference object; a generation unit for inputting the input data to a model to generate output data for specifying the target position, a reference position that is a position of the reference object, and a positional relationship of the target position with respect to the reference position; and an update unit for updating a parameter of the model to reduce a loss obtained by inputting the output data and the correct answer data to a loss function. The loss function is based on at least two errors of a first error between the target position specified by the output data and the target position specified by the correct answer data, a second error between the reference position specified by the output data and the reference position specified by the correct answer data, and a third error between the positional relationship specified by the output data and the positional relationship specified by the correct answer data.