DMOFC: Discrimination Metric-Optimized Feature Compression

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

Abstract

Traditional feature compression methods prioritize reconstructing pixel-level data, often at the expense of discriminative power in machine vision tasks. We present Discrimination Metric-Optimized Feature Compression (DMOFC), a framework that optimizes compression for feature discriminability rather than simple similarity. By incorporating a triplet-based discrimination metric, DMOFC ensures that compressed features maintain their categorical relationships, significantly improving performance in tasks like person re-identification and face verification at low bitrates.

Introduction

As video data production scales exponentially, the primary consumers of visual content are increasingly machine vision algorithms rather than humans. This shift necessitates a move from traditional Video Coding for Humans (VCH) to Video Coding for Machines (VCM). While VCH focuses on minimizing Mean Squared Error (MSE) to satisfy human perception, VCM requires preserving the semantic information necessary for downstream tasks. Existing feature compression techniques often treat feature vectors independently, which can lead to "feature overlap" in high-dimensional space, particularly under heavy compression. DMOFC addresses this by explicitly optimizing for inter-feature relationships.

Methodology

The core innovation of DMOFC is the Discrimination Metric (DIS), which replaces or augments the standard similarity-based loss.

The Discrimination Metric

The DIS metric is based on a triplet loss formulation. For an anchor feature $A$, a positive sample $P$ (same category), and a negative sample $N$ (different category), the objective is to ensure: $$d(A, P) + \alpha < d(A, N)$$ where $d(\cdot)$ is a distance metric and $\alpha$ is a predefined margin. This ensures that even after compression, features from the same class remain closer to each other than to features from different classes.

Conceptual diagram of the discrimination metric (a) and t-SNE visualizations of feature clusters (b-d).

Optimization Strategy

The authors evaluate three optimization objectives:

  1. SIM: Optimized solely for similarity (MSE) between original and compressed features.
  2. DIS: Optimized solely for the discrimination metric.
  3. COMBINED: A weighted sum of SIM and DIS, balancing reconstruction fidelity with discriminative power.

The architecture utilizes a simple two-layer fully-connected (FC) network to isolate the impact of the loss function.

Experimental Results

The framework was evaluated on Face Verification (FV) using the LFW dataset and Person Re-identification (ReID) using the DukeMTMC-reID dataset.

Performance at Low Bitrates

DMOFC demonstrates significant advantages at low bitrates (e.g., 16-64 bytes). In Face Verification, the DIS and COMBINED metrics substantially outperformed the SIM approach as information capacity decreased. For ReID, the COMBINED metric consistently achieved the highest mean Average Precision (mAP) across various capacities.

Visual Analysis

t-SNE visualizations confirm the "overlap" phenomenon. Features compressed using only similarity (SIM) tend to scatter and intermingle (as seen in Figure 6c), whereas those optimized with the DIS metric maintain tight, well-separated clusters (as seen in Figure 6b), preserving the boundaries necessary for accurate classification.

Discussion and Conclusion

The effectiveness of DMOFC reveals a "trade-off point" in feature compression. Above a certain bitrate (e.g., 128 bytes for FV), original features are preserved well enough by MSE alone. Below this point, explicit discriminative constraints become critical. Furthermore, the quality of the original features influences the metric's impact; highly discriminative but non-compact features may require careful balancing to avoid excessive intra-class variance.

Future work will explore extending DMOFC to complex non-linear transforms and broader machine vision tasks such as object detection and semantic segmentation.

Summary updated

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.
BibTeX
@inproceedings{gao2024dmofc,
  author = {Changsheng Gao and Yiheng Jiang and Li Li and Dong Liu and Feng Wu},
  title = {{DMOFC: Discrimination Metric-Optimized Feature Compression}},
  booktitle = {PCS},
  year = {2024},
  doi = {10.1109/pcs60826.2024.10566361},
}
Figure