Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SST-EM: Advanced Metrics for Evaluating Semantic, Spatial and Temporal Aspects in Video Editing

Varun Biyyala, Bharat Chanderprakash Kathuria, Jialu Li, Youshan Zhang

TL;DR

SST-EM addresses the inadequacy of traditional frame-wise metrics for video editing by introducing a multi-stage evaluation framework that combines semantic fidelity, object presence, and temporal coherence. It leverages Vision-Language Models for semantic extraction, Grounding DINO for object detection, an LLM agent for object refinement, and Vision Transformers for temporal consistency, culminating in a weighted final score learned from human judgments. The method is validated on a diverse dataset and shows strong alignment with human evaluations, outperforming CLIP-based approaches in capturing both semantic and temporal aspects. The work provides a practical, robust metric with code availability, enabling standardized benchmarking of video editing models and advancing evaluation practices in the field.

Abstract

Video editing models have advanced significantly, but evaluating their performance remains challenging. Traditional metrics, such as CLIP text and image scores, often fall short: text scores are limited by inadequate training data and hierarchical dependencies, while image scores fail to assess temporal consistency. We present SST-EM (Semantic, Spatial, and Temporal Evaluation Metric), a novel evaluation framework that leverages modern Vision-Language Models (VLMs), Object Detection, and Temporal Consistency checks. SST-EM comprises four components: (1) semantic extraction from frames using a VLM, (2) primary object tracking with Object Detection, (3) focused object refinement via an LLM agent, and (4) temporal consistency assessment using a Vision Transformer (ViT). These components are integrated into a unified metric with weights derived from human evaluations and regression analysis. The name SST-EM reflects its focus on Semantic, Spatial, and Temporal aspects of video evaluation. SST-EM provides a comprehensive evaluation of semantic fidelity and temporal smoothness in video editing. The source code is available in the \textbf{\href{https://github.com/custommetrics-sst/SST_CustomEvaluationMetrics.git}{GitHub Repository}}.

SST-EM: Advanced Metrics for Evaluating Semantic, Spatial and Temporal Aspects in Video Editing

TL;DR

SST-EM addresses the inadequacy of traditional frame-wise metrics for video editing by introducing a multi-stage evaluation framework that combines semantic fidelity, object presence, and temporal coherence. It leverages Vision-Language Models for semantic extraction, Grounding DINO for object detection, an LLM agent for object refinement, and Vision Transformers for temporal consistency, culminating in a weighted final score learned from human judgments. The method is validated on a diverse dataset and shows strong alignment with human evaluations, outperforming CLIP-based approaches in capturing both semantic and temporal aspects. The work provides a practical, robust metric with code availability, enabling standardized benchmarking of video editing models and advancing evaluation practices in the field.

Abstract

Video editing models have advanced significantly, but evaluating their performance remains challenging. Traditional metrics, such as CLIP text and image scores, often fall short: text scores are limited by inadequate training data and hierarchical dependencies, while image scores fail to assess temporal consistency. We present SST-EM (Semantic, Spatial, and Temporal Evaluation Metric), a novel evaluation framework that leverages modern Vision-Language Models (VLMs), Object Detection, and Temporal Consistency checks. SST-EM comprises four components: (1) semantic extraction from frames using a VLM, (2) primary object tracking with Object Detection, (3) focused object refinement via an LLM agent, and (4) temporal consistency assessment using a Vision Transformer (ViT). These components are integrated into a unified metric with weights derived from human evaluations and regression analysis. The name SST-EM reflects its focus on Semantic, Spatial, and Temporal aspects of video evaluation. SST-EM provides a comprehensive evaluation of semantic fidelity and temporal smoothness in video editing. The source code is available in the \textbf{\href{https://github.com/custommetrics-sst/SST_CustomEvaluationMetrics.git}{GitHub Repository}}.
Paper Structure (21 sections, 6 equations, 4 figures, 5 tables)

This paper contains 21 sections, 6 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset Collection
  4. Data Preparation for Evaluation
  5. Methodology
  6. Stage 1: Context Similarity Score
  7. Stage 2: Object Detection Score
  8. Stage 3: Temporal Consistency Score
  9. Stage 4: Final Score Calculation
  10. Optimization and Regression
  11. Finalizing Weights Optimization
  12. Results
  13. Comparison with Other Metrics
  14. Pearson Correlation
  15. Spearman Rank Correlation
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Examples of generated video frames with generation prompt.
  • Figure 2: Comparison between Final and Human Eval Results
  • Figure 3: Pipeline for evaluating video editing models using custom metrics: semantic analysis, object detection, and temporal consistency
  • Figure 4: Comparison of results between different video editing models. The primary object of interest and the original video are highlighted in red.