Frame-Level Video Segmentation for Precise Category Boundaries

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current video classification and tagging methods fail to accurately classify different categories of instances within a video and cannot determine the starting and ending time periods of these instances, leading to inefficiencies in managing large volumes of video content.

Innovation Solution

Utilize an image classifier trained with pre-tagged image data to segment videos by grouping adjacent frames with the same category, employing temporal smoothing techniques like CRF models to refine classifications and determine instance boundaries.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If video classifying and tagging method is used to understand video content, then the entire video can be categorized with one or more category tags, but different categories of instances within the video cannot be classified and their time periods cannot be positioned

Engineering Contradiction:
Improvevideo categorizationVSAvoidinstance category details
Core Design Contradiction:
Extent of automationVSLoss of information

Solution Approach 1:

The patent segments the video into multiple video segments based on image classification results of individual frames. Each video segment corresponds to a specific category instance with defined time boundaries. This segmentation approach enables the system to identify and classify multiple different categories of instances within a single video, preserving detailed instance information that would otherwise be lost in whole-video classification.

Inventive Principle:
Principle #1Segmentation

2Productivity

If current video classifying and tagging method is applied, then video processing can be completed efficiently, but starting and ending time periods of multiple different categories of instances cannot be determined

Engineering Contradiction:
Improvevideo processing efficiencyVSAvoidinstance time boundary detection
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent performs preliminary image classification on individual video frames before assembling them into video segments. By classifying each frame independently first, the system establishes a foundation for precise time boundary detection. The classification results of adjacent frames are then combined to determine the starting and ending time periods of each instance category, achieving both efficiency and precision.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent transitions from whole-video classification to frame-level classification, adding a temporal dimension to the analysis. By examining each frame individually and then grouping adjacent frames with the same category, the system can precisely identify time boundaries of different instance categories while maintaining processing efficiency through automated classification.

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

3Speed

If traditional video understanding methods are used, then processing speed can be maintained, but rich semantic information of objects with multiple semantics cannot be preserved

Engineering Contradiction:
Improvevideo processing speedVSAvoidsemantic information
Core Design Contradiction:
SpeedVSLoss of information

Solution Approach 1:

The patent segments video content into distinct video segments, each representing a specific category instance. This segmentation preserves the semantic information of objects with multiple semantics by assigning each object instance its own dedicated time segment with accurate category labels, preventing semantic information loss while maintaining processing speed through efficient classification algorithms.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3792818B1Video processing method and device, and storage medium
Publication Date: 2026.02.25 BEIJING XIAOMI MOBILE SOFTWARE CO LTD
  • EP3792818B1 patent drawingFigure 1
  • EP3792818B1 patent drawingFigure 2
  • EP3792818B1 patent drawingFigure 3

AI summary

The video processing method includes: collecting a plurality of video frames sequentially from a to-be-processed video (110); for each of the plurality of video frames, classifying an image corresponding to each of the plurality of video frames to obtain a classification result (120); performing a temporal smoothing processing on the classification results for the images corresponding to the video frames, to determine at least one category associated with each of the video frames (130); and grouping the adjacent video frames of a same category into a video segment, to obtain a plurality of video segments (140). If a same video frame belongs to different categories, the video frame is categorized into different video segments of the categories corresponding to the video frame.