Object Detection Beyond Fixed Categories Using Language Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems rely heavily on pre-defined text categories, leading to missed and false detections when text is lacking or incomplete, affecting accuracy.

Innovation Solution

An object detection method using an object detection model to extract visual feature representations and a language model to generate text sequences indicating object categories, without relying on pre-defined text, leveraging the language model's semantic analysis capabilities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If pre-defined text categories are used for object detection, then the detection process is simple and fast, but the detection accuracy decreases when text is lacking or incomplete

Engineering Contradiction:
Improvedetection speedVSAvoiddetection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent introduces text embedding vectors as an intermediary between visual features and category predictions. These embeddings capture semantic information from text descriptions and serve as a bridge, allowing the model to leverage both visual and textual information for more accurate detection while maintaining computational efficiency through the established text-to-visual attention mechanism

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent combines visual features and text embedding vectors into a composite representation that integrates both modalities. This composite approach allows the model to benefit from the complementary strengths of visual and textual information, improving detection accuracy for objects that may be ambiguous in one modality alone while maintaining the efficiency of the unified processing framework

Inventive Principle:
Principle #40Composite materials

2Device complexity

If pre-defined text categories are used for object detection, then the system structure is simple, but the adaptability to new categories decreases

Engineering Contradiction:
Improvesystem structureVSAvoidcategory adaptability
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent implements a dynamic attention mechanism where text embeddings are computed on-the-fly from input text descriptions rather than being fixed to pre-defined categories. This allows the system to adapt to new object categories and descriptions without requiring retraining of the entire model, as the text embedding layer can accommodate any textual input while the visual feature extraction remains efficient

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The text embedding layer serves multiple functions: it can process predefined category names, free-text descriptions, and multi-lingual inputs uniformly. This universal text processing capability allows the same model architecture to handle diverse categorization schemes and adapt to new categories through text descriptions alone, eliminating the need for separate processing pipelines for different category systems

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4597440A1Method, apparatus, device, and storage medium for object detection
Publication Date: 2025.08.06 BEIJING YOUZHUJU NETWORK TECH CO LTD
  • EP4597440A1 patent drawingFigure 1~2B
  • EP4597440A1 patent drawingFigure 3~4B
  • EP4597440A1 patent drawingFigure 5~6

AI summary

Provided in the disclosure are a method, an apparatus, a device, and a storage medium for object detection. The method includes: extracting, by using an object detection model, a group of visual feature representations of a target image, the group of visual feature representations including respective visual feature representations of at least one object area in the target image; and generating, by using a language model, a group of text sequences based on the group of visual feature representations, each text sequence indicating at least one category to which an object in an object area corresponding to the visual feature representation belongs.