Weakly-Supervised Object Detection with Pseudo-Label Refinement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Creating labeled training data for machine learning models is a long and complicated process, especially for object detection tasks, which hinders the efficiency of object identification and analysis.

Innovation Solution

A framework for weakly-supervised object detection using a multi-branch neural network that accepts image-level tags without spatial information, employing refinement branches to improve detection accuracy through pseudo-labels and a drop-block approach to address instance ambiguity and memory constraints.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional supervised object detection is used, then detection accuracy is improved, but the complexity and time required for creating labeled training data increases significantly

Engineering Contradiction:
Improvedetection accuracyVSAvoidtraining data preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system uses the detected images themselves to generate training data automatically through the multi-branch neural network framework. The detection results are fed back to create pseudo-labels and refine the model, eliminating the need for manual annotation of training data while maintaining detection accuracy

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent introduces an intermediate multi-branch neural network framework that bridges the gap between weakly-labeled data and accurate detection. This framework includes branches for feature extraction, pseudo-label generation, and refinement, acting as a mediator to transform minimal labeled data into effective training signals

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If traditional supervised object detection is used, then detection accuracy is improved, but the time required for training and processing increases

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

Solution Approach 1:

The system performs preliminary feature extraction and representation learning using the multi-branch network before final detection. By pre-processing and extracting robust features in advance, the training process is accelerated while maintaining or improving detection accuracy

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The training process is segmented into multiple branches that operate in parallel: one branch for feature extraction, another for pseudo-label generation, and a third for refinement. This segmentation allows simultaneous processing of multiple tasks, reducing overall training time while achieving accurate detection

Inventive Principle:
Principle #1Segmentation

3Measurement precision

If detailed spatial labeling is performed, then detection precision is improved, but memory consumption increases

Engineering Contradiction:
Improvedetection precisionVSAvoidmemory consumption
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential spatial information needed for detection from the labeled data, removing redundant detailed annotations. By taking out only the critical spatial features rather than storing complete detailed labels, memory consumption is reduced while detection precision is maintained

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system transforms detailed spatial labeling into a different dimensional representation using the multi-branch network. Instead of storing extensive spatial coordinates, the labels are transformed into feature space representations that consume less memory while preserving detection precision through the network's learned features

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

Data Source

PatentEP3754560B1Weakly-supervised object detection using one or more neural networks
Publication Date: 2025.10.15 NVIDIA CORP
  • EP3754560B1 patent drawingFigure 1A~1C
  • EP3754560B1 patent drawingFigure 2
  • EP3754560B1 patent drawingFigure 3A~3D

AI summary

Apparatuses, systems, and techniques to detect object in images including digital representations of those objects. In at least one embodiment, one or more objects are detected in an image based, at least in part, on one or more pseudo-labels corresponding to said one or more objects.