Video Feature Extractor Temporal Prediction Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current unsupervised video feature extraction methods fail to effectively capture and utilize temporal information from video sequences, leading to suboptimal performance in downstream applications such as action recognition and behavior prediction, due to the lack of strong training signals that encourage the model to explore temporal structures.

Innovation Solution

The proposed method employs a self-supervised learning approach by using joint prediction of past and future subsequences as a pretext task, where the video feature extractor is trained to estimate the joint predictability of preceding and succeeding frames, thereby encouraging the extraction of high-quality features that emphasize temporal aspects. This is achieved through a contrastive loss function that compares feature representations across different temporal orders, using both easy and hard negative examples to force the model to distinguish between correct and swapped temporal orders.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If unsupervised learning is used for video feature extraction, then annotation cost and error are reduced, but the ability to capture temporal information deteriorates

Engineering Contradiction:
Improveannotation costVSAvoidtemporal information
Core Design Contradiction:
Ease of manufactureVSLoss of information

Solution Approach 1:

The system uses self-supervised learning where the model generates its own training signals by predicting future frames from past frames. This eliminates the need for external annotations while still providing strong supervisory signals that capture temporal dynamics through the prediction task itself.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The model is trained to predict future frames before actual supervision is needed. By performing the prediction task as a preliminary training objective, the model learns temporal relationships in advance, which then transfers to downstream tasks without requiring labeled temporal data.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If supervised learning with action labels is used, then labeled data provides supervision signal, but the supervision signal is weak and does not force exploration of temporal information

Engineering Contradiction:
Improvesupervision signalVSAvoidtemporal information
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

Instead of using action labels to supervise feature extraction, the approach inverts the problem: it uses the feature extractor to predict future frames. This inversion creates a much stronger supervisory signal because the model must learn comprehensive temporal representations to accurately predict future states, rather than merely classifying actions.

Inventive Principle:
Principle #13The other way round (Inversion)

3Measurement precision

If dense predictive coding with NCE loss is used, then future representation prediction is optimized, but the model still fails to sufficiently explore temporal information

Engineering Contradiction:
Improveprediction accuracyVSAvoidtemporal information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The approach merges multiple temporal perspectives by training the model to predict both immediate future frames and longer-term future representations simultaneously. This combination of short-term and long-term prediction tasks ensures comprehensive exploration of temporal information at multiple scales.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3989106B1Unsupervised training of a video feature extractor
Publication Date: 2024.05.15 ROBERT BOSCH GMBH
  • EP3989106B1 patent drawingFigure 1~2
  • EP3989106B1 patent drawingFigure 3
  • EP3989106B1 patent drawingFigure 4

AI summary

The invention relates to a computer-implemented unsupervised learning method (600) of training a video feature extractor. The video feature extractor is configured to extract a feature representation from a video sequence. The method uses training data representing multiple training video sequences. From a training video sequence of the multiple training video sequences, a current subsequence; a preceding subsequence preceding the current subsequence; and a succeeding subsequence succeeding the current subsequence are selected. The video feature extractor is applied to the current subsequence to extract a current feature representation of the current subsequence. A training signal is derived from a joint predictability of the preceding and succeeding subsequences given the current feature representation. The parameters of the video feature extractor are updated based on the training signal.