E2E Learnable Video Feature Compression for Object Detection and Tracking

다국어 초록 (Multilingual Abstract)

This paper proposes a novel video feature map compression method that effectively reduces temporal redundancy between sequential frames using a learnable entropy model. The proposed model leverages the previously decoded single-scale feature map to improve compression efficiency without explicitly estimating motion information. The proposed model applies the previously decoded single-scale feature map in the Feature Reduction, Feature Restoration, and Feature Inner Coding stages, significantly improving compression performance. Experimental results show that the method achieves superior performance compared to frame-wise compression models, with an average BD-rate gain of 64.23%, and outperforms the current state-of-the-art FCTM-v4.0 by 4.82%. These results confirm that the proposed method effectively compresses video feature maps while maintaining general performance across various machine vision tasks, offering a promising approach to video feature compression.

목차 (Table of Contents)

• 1. Introduction 1
• 2. Related Works 3
• 1) MPEG FCM 3
• 2) L-MSFC-v2 5
• 3) Learned Video Compression 8
• 3. Methodology 10
• 1) Overview 10
• 2) Single-Scale Feature Fusion in Feature Reduction 12
• 3) Single-Scale Feature Compensation in Feature Restoration 14
• 4) Temporal-based Entropy Model in Feature Inner Coding 14
• 5) Training Loss 16
• 4. Experiments 18
• 1) Training and Testing Details 18
• 2) Experimental Results 19
• 2-1) Evaluation Conditions 19
• 2-2) Performance 19
• 3) Ablation Study 23
• 3-1) Simplified Architecture of the Proposed Model 23
• 3-2) Comparison with FCTM-v4.0 [13] 23
• 5. Conclusion 25
• 6. References 25