Feature Compression for Rate Constrained Object Detection on the Edge

Zhongzheng Yuan, Samyak Rawlekar, Siddharth Garg, Elza Erkip, Yao Wang

Abstract

Deploying state-of-the-art object detection on resource-constrained edge devices necessitates a balance between local computation, network bandwidth, and detection accuracy. This paper presents a split-computation framework for YOLOv5 that utilizes a learnable feature compression module. By optimizing for detection-specific loss rather than traditional image reconstruction metrics, the proposed system achieves superior mean Average Precision (mAP) at low bitrates. Furthermore, the lightweight design enables real-time performance on mobile CPUs, significantly reducing end-to-end latency compared to standard image-compression offloading baselines.

1. Introduction

Computer vision applications on mobile platforms—such as drones, robotics, and wearable devices—are often hindered by a "triple threat" of constraints: limited local computational power, restricted battery life, and fluctuating network bandwidth. Traditional deployment strategies typically fall into two categories: local execution, which often exceeds the device's thermal and power envelopes, or full offloading. In full offloading, the device compresses and transmits raw images (using codecs like JPEG or BPG) to a powerful server. However, these codecs are designed for human perception and often discard high-frequency details essential for neural network performance or waste bits on background textures irrelevant to the task.

This research explores the "split computation" paradigm. By splitting the YOLOv5 architecture between the edge device and the server, the system can compress intermediate feature maps. This approach aims to minimize both the data transmitted and the mobile-side computational load without compromising detection accuracy.

2. Proposed Methodology

The core of the framework is a learnable feature compression module integrated into the YOLOv5 backbone. The architecture is defined by several key design choices:

2.1 Split Point Selection

The researchers evaluated splitting the network at the 3rd (D3) and 4th (D4) downsampling layers. While splitting earlier reduces the computational burden on the mobile device, splitting later produces more spatially compact and sparse features, which are inherently easier to compress.

2.2 Channel Reduction and Hyperprior Compression

To address inter-channel redundancy, a $1 \times 1$ convolution layer reduces the feature channels (e.g., from 128 to 4) before transmission. These reduced features are then processed by a hyperprior-based compression module. This module uses an encoder-decoder pair to learn the probability distribution of the features, allowing an Arithmetic Coder to achieve high compression efficiency.

2.3 Detection-Aware Training

Unlike standard learned image compression that minimizes Mean Squared Error (MSE), this system is trained end-to-end using the YOLO detection loss. This ensures the compressor prioritizes information critical for identifying objectness, class labels, and bounding box coordinates. To ensure stability, the authors employed a two-stage training strategy: first training the channel reduction layers without quantization, then introducing the compression bottleneck.

3. Experimental Setup

The framework was evaluated on two primary benchmarks:

  1. COCO Dataset: The standard benchmark for general object detection featuring 80 classes.
  2. COCO-Traffic: A specialized subset containing 9 classes relevant to autonomous navigation (e.g., Person, Car, Bus, Traffic Light). This subset demonstrates the advantages of task-specific compression in specialized environments.

4. Results and Discussion

4.1 Rate-Accuracy Performance

On the COCO dataset, the proposed method consistently outperformed BPG and other learned image compression baselines in the low-to-medium bitrate range. The gains were even more significant on the COCO-Traffic dataset, where the model learned to ignore features irrelevant to the traffic domain, maintaining high mAP at extremely low bitrates where traditional methods failed.

4.2 Latency and Efficiency

A critical finding of the study is the framework's suitability for mobile hardware. While many learned compressors are too computationally intensive for edge deployment, this feature compressor is highly optimized:

  • Mobile Latency: Feature compression took approximately 63ms on a mobile CPU, compared to 415ms for state-of-the-art learned image codecs.
  • End-to-End Speedup: The total latency (mobile processing + transmission + server processing) was reduced by 6.06x compared to learned image compression and 3.93x compared to BPG offloading.

5. Limitations and Future Work

Despite the performance gains, the authors noted that splitting at deeper layers (D4) can impact the detection of very small objects due to the loss of high-resolution spatial information and the removal of specific skip connections. Future research will focus on adapting this methodology to Transformer-based architectures and improving robustness against varying network conditions, such as packet loss and jitter.

6. Conclusion

This work demonstrates that task-specific feature compression is a viable and efficient alternative to traditional image offloading for edge-based object detection. By shifting the optimization goal from visual fidelity to detection accuracy, the proposed framework achieves a superior trade-off between bitrate, latency, and precision, enabling real-time computer vision on resource-constrained devices.

Summary updated

BibTeX
@inproceedings{yuan2022feature,
  author = {Zhongzheng Yuan and Samyak Rawlekar and Siddharth Garg and Elza Erkip and Yao Wang},
  title = {{Feature Compression for Rate Constrained Object Detection on the Edge}},
  booktitle = {MIPR},
  year = {2022},
  doi = {10.1109/mipr54900.2022.00008},
}
Figure