Teacher-Student Object Detection With Adaptive Pseudo-Labeling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing semi-supervised learning techniques face challenges in transferring from image classification to object detection due to complexities in architecture design, necessitating manual annotation of large datasets for object detection, which is tedious and resource-intensive.

Innovation Solution

A neural network framework utilizing a teacher-student approach with pseudo-labeling, exponential moving average updates, and adaptive loss functions to leverage unlabeled data for training a student neural network with limited labeled data, incorporating jitter-bagging and adaptive threshold mechanisms for improved bounding box predictions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If semi-supervised learning techniques are applied to object detection, then the need for labeled data is reduced, but the architecture design complexity increases

Engineering Contradiction:
Improvelabeled dataVSAvoidarchitecture design
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

The object detection architecture is segmented into two distinct networks: a teacher network that generates pseudo-labels and a student network that learns from them. This segmentation allows the system to leverage unlabeled data while maintaining a manageable architecture by distributing different functions across separate networks rather than creating a single complex unified system.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The teacher network performs preliminary action by generating pseudo-labels for unlabeled data before the student network uses them for training. This preliminary labeling step enables the student network to learn from both labeled and unlabeled data without requiring complex architecture modifications, as the pseudo-labels are prepared in advance.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If manual annotation is performed for object detection, then the training data quality is improved, but the resource consumption and time required increase

Engineering Contradiction:
Improvedata qualityVSAvoidannotation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system implements self-service through automated pseudo-label generation. The teacher network automatically generates labels for unlabeled data without human intervention, enabling the system to scale to large datasets without proportionally increasing annotation time. This self-labeling capability maintains data quality while dramatically reducing the time and resources required compared to manual annotation.

Inventive Principle:
Principle #25Self-service

3Adaptability or versatility

If a teacher-student network architecture is used, then the utilization of unlabeled data is improved, but the computational resources required increase

Engineering Contradiction:
Improveunlabeled data utilizationVSAvoidcomputational resources
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The training process implements continuity of useful action by simultaneously processing both labeled and unlabeled data through the teacher-student framework. The teacher network continuously generates pseudo-labels while the student network continuously learns from both labeled and pseudo-labeled data, maximizing the utilization of all available data without requiring separate processing stages that would waste computational resources.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4463830B1Teaching student network for end-to-end semi-supervised object detection
Publication Date: 2026.04.01 SONY GROUP CORP
  • EP4463830B1 patent drawingFigure 1
  • EP4463830B1 patent drawingFigure 2
  • EP4463830B1 patent drawingFigure 3

AI summary

A system and method for end-to-end semi-supervised object detection is provided. The system retrieves labeled and unlabeled images from an image dataset and generates an input batch by application of image transformation(s) on the images. The system further generates a first result for each image of the input batch by application of a teacher neural network on the input batch. For an object in an unlabeled image of the batch, the first result includes candidate bounding boxes and scores for the boxes. The system determines a threshold score based on the scores and selects a foreground bounding box from the candidates. The system generates a second result by application of a student neural network on the unlabeled image and computes a training loss over the input batch based on the foreground bounding box and the second result. The system trains the student neural network based on the training loss.