Efficient Object Detection in Compressed Domain by Exploiting Knowledge Distillation from Pixel Domain

Area of Science:

• Computer Vision
• Machine Learning
• Video Processing

Background:

• High-definition video data requires efficient real-time edge analytics.
• Traditional object detection relies on pixel-domain inputs, creating latency bottlenecks due to decoding.
• Compressed-domain processing offers a potential solution for faster analytics.

Purpose of the Study:

• To develop a fast and efficient object detection framework operating directly on partially decoded compressed-domain video data.
• To overcome the latency bottleneck associated with traditional pixel-domain object detection pipelines.
• To improve the efficiency and accuracy of object detection for edge analytics.

Main Methods:

• Proposed a dual-phase framework for compressed-domain object detection.
• Introduced Low-Frequency Spectral Prioritization for partial decoding, reducing data and accelerating processing.
• Employed multi-granularity cross-domain knowledge distillation to transfer knowledge from a pixel-domain teacher to a compressed-domain student network.

Main Results:

• The proposed framework enables object detection directly on partially decoded HEVC (High Efficiency Video Coding) data.
• Experiments on COCO-mini dataset with RetinaNet, FCOS, and GFL showed improved mAP (+0.96%) compared to pixel-domain baselines.
• Achieved reduced average decoding latency by retaining fundamental residual coefficients.

Conclusions:

• The novel dual-phase framework offers a viable solution for efficient, low-latency object detection in compressed video streams.
• Compressed-domain processing, combined with knowledge distillation, effectively mitigates traditional decoding bottlenecks.
• This approach enhances the feasibility of real-time edge analytics for high-definition video data.