Encoder Model Training With Temporal Offset Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing encoder models struggle to effectively extract features from sequences of frames, particularly in video data, as they often lose important temporal information due to the lack of robustness in handling temporal shifts, which is crucial for tasks like fine-grained human action recognition.

Innovation Solution

A computer-implemented method for training an encoder model by obtaining subsequences from an input signal, encoding them, and offsetting based on shifts between the sequence and subsequence, allowing the model to learn features that are aware of the succession dimension, thereby improving generalization and accuracy in downstream tasks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Stability of the object's composition

If feature persistence is enforced over the sequence, then temporal invariance is improved, but temporal order information is lost

Engineering Contradiction:
Improvefeature persistenceVSAvoidtemporal order information
Core Design Contradiction:
Stability of the object's compositionVSLoss of information

Solution Approach 1:

The patent segments the sequence into multiple subsequences with different temporal offsets. Each subsequence is encoded separately, allowing the model to capture both temporal invariance (through shared encoding) and temporal order information (through offset differentiation). This segmentation resolves the contradiction by processing the sequence in differentiated temporal segments rather than as a single persistent unit.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by assigning different temporal offsets to different subsequences. Instead of treating the entire sequence uniformly, each subsequence receives a customized offset that preserves its temporal position information. This allows feature persistence to be maintained at the subsequence level while temporal order information is preserved through the offset assignments.

Inventive Principle:
Principle #3Local quality

2Reliability

If random spatial cropping is used for data augmentation, then training robustness is improved, but temporal information is insufficient

Engineering Contradiction:
Improvetraining robustnessVSAvoidtemporal information
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent extends data augmentation from purely spatial operations to spatio-temporal operations. By introducing temporal offsets as an additional dimension of transformation, the model learns to handle both spatial variations (through cropping) and temporal variations (through offsets). This dimensional extension resolves the contradiction by adding temporal awareness to the augmentation process.

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

Solution Approach 2:

The patent changes the augmentation parameters from purely spatial (crop size, crop position) to spatio-temporal (crop size, crop position, temporal offset). This parameter extension allows the model to generate training samples that preserve temporal information while maintaining robustness through varied spatial and temporal transformations.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4411671A1Method and system for training an encoder model
Publication Date: 2024.08.07 TOYOTA JIDOSHA KK
  • EP4411671A1 patent drawingFigure 1~2
  • EP4411671A1 patent drawingFigure 3
  • EP4411671A1 patent drawingFigure 4(a)~4(b)

AI summary

A computer-implemented method for training an encoder model (f) to extract features from an input signal comprising a sequence of frames, the computer-implemented method comprising: - obtaining at least one subsequence (20) from at least one such input signal; - encoding (22) the at least one subsequence with the encoder model (f); - offsetting the encoded at least one subsequence (24) based on a shift between the sequence of the at least one input signal and the at least one subsequence; - training (30) the encoder model (f) based on the offset encoded at least one subsequence (26).