Efficient One-Stage Detectors for Real-Time Edge Applications

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2477 papers540 conceptsfrozen 2026-10-04

In the laboratory, computer vision models often enjoy the luxury of high-end GPUs and unlimited power. But in the real world—on a drone inspecting power lines, a wearable device for the visually impaired, or a sensor in a coal mine—intelligence must be squeezed into low-power "edge" hardware. The research direction emerging from these 241 papers focuses on bridging this gap: transforming heavy, computationally expensive object detection architectures into lean, high-speed engines capable of real-time inference on the edge.

The Shift to Single-Stage Architectures

The fundamental intuition behind this research is a move from "accuracy at any cost" to "sufficient accuracy at high speed." Traditionally, object detection was dominated by two-stage models like Faster R-CNN, which first propose regions of interest and then classify them. While precise, these are often too slow for real-time use on mobile hardware.

As Xiren Miao and Hao Jiang note in their review of power line inspection, the field has shifted toward single-stage detectors like the YOLO (You Only Look Once) family and SSD (Single Shot MultiBox Detector). These models treat detection as a single regression problem, mapping straight from image pixels to bounding box coordinates and class probabilities in one forward pass. This architectural choice is the primary driver for achieving the low latency required for applications like autonomous drone flight and real-time safety monitoring.

Architectural Slimming and Feature Enhancement

Simply choosing a one-stage model isn't enough for the most constrained environments. Researchers are actively "slimming" these models while adding surgical enhancements to maintain precision.

A recurring strategy is the use of lightweight backbones—the part of the network that extracts features. For instance, Ruina Du and colleagues developed CIB-SE-YOLOv8, which replaces standard blocks with Compact Inverted Blocks (CIB). This reduces the parameter count and computational load (GFLOPs) without losing the ability to detect safety equipment on construction sites. Similarly, Xiaojun Li and Miao Li introduced DSD-YOLOv8, which uses deformable convolutions to help the model adapt to the irregular shapes of drill pipes in low-visibility coal mines.

To compensate for the reduced depth of these lightweight models, researchers integrate attention mechanisms. These modules act like a spotlight, telling the network which parts of the image are most important. Mingjie Jiang and Xinqi Fan utilized a Context Attention Module in RetinaFaceMask to ensure that even a small, mobile-friendly model could accurately detect whether a face mask was being worn correctly.

Hardware-Aware Optimization

The ultimate goal of this research is deployment on specific, often underpowered, hardware. This requires a "hardware-aware" design philosophy.

  • Raspberry Pi and Mobile CPUs: For wearable assistive devices, Lucas Nadolskis demonstrated that mobile-optimized single-shot detectors can provide real-time auditory feedback for social distancing.
  • NVIDIA Jetson Nano/Xavier: For more intensive tasks like tracking active shooters, Jiale Feng and team showed that even complex tracking pipelines can achieve functional frame rates on the Jetson Nano by leveraging synthetic data and optimized pipelines.
  • ARM Processors: Even the traditionally "heavy" two-stage detectors are being reimagined for these chips. Zheng Qin and Zeming Li developed ThunderNet, the first two-stage detector to achieve real-time performance on ARM processors by radically thinning the detection "head."

Challenging the Speed-Accuracy Trade-off

The prevailing wisdom in computer vision has long been that you must choose between speed and accuracy. However, this research direction suggests that for specialized tasks, this trade-off is a false dichotomy.

By focusing on specific domains—such as night-time UAV detection or agricultural weed identification—researchers can optimize architectures for the specific features of those environments. Zeming Li and Chao Peng proved this with Light-Head R-CNN, which actually outperformed many one-stage models in both speed and precision by slimming down the redundant parts of the network.

The result is a new generation of "industrial-strength" vision: models that are small enough to live on a chip, fast enough to react to the world in milliseconds, and accurate enough to be trusted with human safety and infrastructure.

Papers in this version

Showing 60 of 477 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 560 papers
  2. Version 1 · 2026-10-04 · 477 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_single_stage_object_detection_72ad7798,
  title = {single-stage object detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-stage-object-detection/rdr_72ad7798dc994f6883c2921ec23d7827},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_efficient_one_stage_detectors_for_real_t_72ad7798_v1,
  title = {Efficient One-Stage Detectors for Real-Time Edge Applications},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/efficient-one-stage-detectors-for-real-time-edge-applications/rdr_72ad7798dc994f6883c2921ec23d7827/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}