Pixel-wise Actor Localization via Dynamic Filters

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional object localization systems struggle to accurately localize actors and their actions in a video sequence based on a natural language query, particularly when multiple objects with similar appearances are present, as they are limited to single-frame localization and fail to distinguish between different actions performed by these objects.

Innovation Solution

A method and apparatus for pixel-wise localization of actors and actions in a sequence of frames using a fully convolutional neural network that integrates visual and language information, generating dynamic filters based on natural language queries to perform segmentation and localization, enabling fine-grained distinction between actors and their actions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional object localization systems use single-frame localization, then the system complexity is reduced, but the measurement precision of action distinction deteriorates

Engineering Contradiction:
Improveaction distinction precisionVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the video sequence into multiple frames and processes each frame independently through the convolutional neural network. This allows the system to maintain computational efficiency while achieving precise action distinction by analyzing temporal patterns across segmented frames rather than processing the entire video as one complex unit.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from single-frame (2D spatial) localization to multi-frame sequence analysis (adding the temporal dimension). By processing sequences of frames and analyzing changes over time, the system achieves action distinction precision without requiring a completely new computational approach, thus managing complexity effectively.

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

2Reliability

If conventional systems rely on single-frame localization, then the processing time is reduced, but the reliability of actor identification deteriorates

Engineering Contradiction:
Improveactor identification reliabilityVSAvoidprocessing time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent extracts visual features and performs preliminary processing on each frame before assembling the final action localization result. By preparing frame data in advance and using efficient convolutional neural network architectures, the system achieves reliable multi-frame analysis without excessive processing time delays.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent maintains continuous processing of video frames through the convolutional neural network, ensuring that action localization is performed consistently across all frames. This continuous approach improves reliability by eliminating gaps in analysis while keeping processing time manageable through optimized computational pipelines.

Inventive Principle:
Principle #20Continuity of useful action

3Manufacturing precision

If conventional localization uses bounding boxes for actors, then the ease of operation is improved, but the manufacturing precision of pixel-wise localization deteriorates

Engineering Contradiction:
Improvepixel-wise localization precisionVSAvoidease of operation
Core Design Contradiction:
Manufacturing precisionVSEase of operation

Solution Approach 1:

The patent segments the actor representation from coarse bounding boxes to fine pixel-level segments. By dividing the actor region into multiple pixel-level segments and processing each through the convolutional neural network, the system achieves precise pixel-wise localization while maintaining operational simplicity through automated processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different processing qualities to different regions of the image. Instead of uniform processing, the system focuses computational resources on specific pixel regions where actors are detected, applying high-precision convolutional neural network analysis only where needed. This local quality approach improves pixel-wise localization precision without proportionally increasing overall computational complexity.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3711022B1Spatio-temporal action and actor localization
Publication Date: 2025.08.20 QUALCOMM INC
  • EP3711022B1 patent drawingFigure 1
  • EP3711022B1 patent drawingFigure 2A~2C
  • EP3711022B1 patent drawingFigure 2D

AI summary

A method of pixel-wise localization of an actor and an action in a sequence of frames includes receiving a natural language query describing the action and the actor. The method also includes receiving the sequence of frames. The method further includes localizing the action and the actor in the sequence of frames based on the natural language query.