Detection Transformer Query Partitioning for Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current object detection techniques, such as the DETR model, do not fully utilize all object queries during the inference phase, leading to suboptimal performance in digital image detection tasks.

Innovation Solution

The proposed solution involves a detection transformer model with a self-attention layer that allows independent learning for each subgroup of object requests. This model uses a mask of attention to restrict information transmission within subgroups and employs a cost function to optimize query diversification, leveraging both positive and negative requests during the learning phase.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If all object queries are used during inference phase, then detection accuracy improves, but computational complexity and processing time increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent partitions the set of object queries into multiple subgroups during the inference phase. Each subgroup processes a portion of the queries independently, allowing the system to utilize all queries for improved detection accuracy while managing computational complexity through structured division. This segmentation enables parallel processing and reduces the burden on any single computational unit.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If all object queries are utilized during inference, then detection performance improves, but processing time increases

Engineering Contradiction:
Improvedetection performanceVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

By dividing object queries into subgroups that can be processed in parallel, the patent reduces the sequential processing time while maintaining utilization of all queries. The structured partitioning allows simultaneous evaluation of multiple query subsets, thereby improving detection performance without linearly increasing total processing time.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs query partitioning and organization during the inference phase before full processing begins. This preliminary structuring of queries into manageable subgroups enables more efficient subsequent processing, reducing the overall time required to evaluate all queries while maintaining comprehensive detection performance.

Inventive Principle:
Principle #10Preliminary action

3Device complexity

If traditional DETR model uses single group of queries, then model simplicity is maintained, but information utilization is suboptimal

Engineering Contradiction:
Improvemodel simplicityVSAvoidinformation utilization
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent introduces a partitioning mechanism that divides the query set into multiple subgroups, enabling the model to capture diverse object characteristics and relationships that would be missed by a single-group approach. This segmentation preserves model simplicity through structured organization while significantly improving information utilization during detection.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4535297A1Detection transformer for object detection in digital images based on a group partitioning of the set of object requests
Publication Date: 2025.04.09 PARI MUTUEL URBAIN
  • EP4535297A1 patent drawingFigure 1~2
  • EP4535297A1 patent drawingFigure 3
  • EP4535297A1 patent drawingFigure 4

AI summary

Method for learning an object detection model in a digital image, said model being of the detection transformer type and comprising a first neural network (N1) adapted to determine features within the image, an encoder (NE) to generate information from these features, and a decoder comprising a self-attention layer (SL), the decoder being adapted to generate predictions (P1, P2... PN) from object queries (Q1, Q2... QN) based on information (Eo) provided by the encoder (NE), in which the queries are structured as subgroups (G1, G2... GK), the learning comprising a search for matching the predictions of each subgroup with the same training set.