Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DMOFC: Discrimination Metric-Optimized Feature Compression

Changsheng Gao, Yiheng Jiang, Li Li, Dong Liu, Feng Wu

TL;DR

This work addresses the gap in feature compression for machines by incorporating inter-feature discriminability through a discrimination metric. The authors define a triplet-based metric $DIS(A,P,N) = \max(d(A,P)-d(A,N)+\alpha,0)$ with $d$ as the MSE distance and $\alpha = 0.3$, and optimize a simple two-FC feature compressor to preserve class separation at varying information capacities. Experiments on face verification and person re-identification show that the discrimination metric improves performance at low information capacity, while at high capacity traditional similarity-based optimization can prevail; a combined objective often yields robust results across tasks and capacities. The study also explores how the discriminability of the original features influences the effectiveness of the discrimination metric, revealing a trade-off between preserving original discriminability and enforcing inter-class separation. This work highlights the practical potential of discriminability-aware feature compression for machine analysis tasks and suggests directions for extending the framework to more vision tasks and inter-feature relationships.

Abstract

Feature compression, as an important branch of video coding for machines (VCM), has attracted significant attention and exploration. However, the existing methods mainly focus on intra-feature similarity, such as the Mean Squared Error (MSE) between the reconstructed and original features, while neglecting the importance of inter-feature relationships. In this paper, we analyze the inter-feature relationships, focusing on feature discriminability in machine vision and underscoring its significance in feature compression. To maintain the feature discriminability of reconstructed features, we introduce a discrimination metric for feature compression. The discrimination metric is designed to ensure that the distance between features of the same category is smaller than the distance between features of different categories. Furthermore, we explore the relationship between the discrimination metric and the discriminability of the original features. Experimental results confirm the effectiveness of the proposed discrimination metric and reveal there exists a trade-off between the discrimination metric and the discriminability of the original features.

DMOFC: Discrimination Metric-Optimized Feature Compression

TL;DR

This work addresses the gap in feature compression for machines by incorporating inter-feature discriminability through a discrimination metric. The authors define a triplet-based metric with as the MSE distance and , and optimize a simple two-FC feature compressor to preserve class separation at varying information capacities. Experiments on face verification and person re-identification show that the discrimination metric improves performance at low information capacity, while at high capacity traditional similarity-based optimization can prevail; a combined objective often yields robust results across tasks and capacities. The study also explores how the discriminability of the original features influences the effectiveness of the discrimination metric, revealing a trade-off between preserving original discriminability and enforcing inter-class separation. This work highlights the practical potential of discriminability-aware feature compression for machine analysis tasks and suggests directions for extending the framework to more vision tasks and inter-feature relationships.

Abstract

Feature compression, as an important branch of video coding for machines (VCM), has attracted significant attention and exploration. However, the existing methods mainly focus on intra-feature similarity, such as the Mean Squared Error (MSE) between the reconstructed and original features, while neglecting the importance of inter-feature relationships. In this paper, we analyze the inter-feature relationships, focusing on feature discriminability in machine vision and underscoring its significance in feature compression. To maintain the feature discriminability of reconstructed features, we introduce a discrimination metric for feature compression. The discrimination metric is designed to ensure that the distance between features of the same category is smaller than the distance between features of different categories. Furthermore, we explore the relationship between the discrimination metric and the discriminability of the original features. Experimental results confirm the effectiveness of the proposed discrimination metric and reveal there exists a trade-off between the discrimination metric and the discriminability of the original features.
Paper Structure (19 sections, 1 equation, 4 figures)

This paper contains 19 sections, 1 equation, 4 figures.

Table of Contents

  1. Introduction
  2. Method
  3. Quality Assessment Analyses
  4. Discrimination Metric
  5. Feature Compression
  6. Optimization Strategy
  7. Experiments
  8. Experimental Settings
  9. Dataset
  10. Training Configuration
  11. Performance Analyses
  12. Face Verification
  13. Person ReID
  14. Feature Analyses
  15. Ablation on Discriminability of Original Features
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Illustration of the discrimination metric and its comparison with the similarity metric. (a) The illustration of the discrimination metric. (b) An exemplary visualization of original features. (c) An exemplary visualization of reconstructed features optimized by the similarity metric. (d) An exemplary visualization of reconstructed features optimized by the discrimination metric.
  • Figure 2: Evaluation performance comparisons between different metrics. (a) Face verification task. (b) Person re-identification task.
  • Figure 3: t-SNE visualization of reconstructed features for different metrics.
  • Figure 4: Evaluation performance comparisons on different metrics. (a) Performance on SimFeat. (b) Performance on DisFeat.