The "gold standard" for object detection has historically been defined by massive, power-hungry models like the R-CNN family or high-resolution Transformers. While these models excel on desktop GPUs, they fail immediately when deployed on the "edge"—the world of 33-gram nano-quadrotors, underwater sensors, and battery-powered industrial cameras.
On these devices, the bottleneck isn't just accuracy; it is a strict "speed tax" imposed by limited thermal budgets, memory footprints, and battery life. Research in this field focuses on closing the gap between high-fidelity computer vision and the harsh constraints of embedded systems.
Architectural Slimming and Feature Reuse
The most direct way to make a model faster is to remove its heaviest components. Standard detectors often rely on fully-connected layers that consume massive amounts of memory. Subarna Tripathi and colleagues demonstrated in LCDet that by swapping these for convolutional layers and implementing 8-bit quantization, one can achieve a 20x performance boost with negligible accuracy loss.
Other researchers look for "free" information already present in the network. Guiying Li and the team behind Relief R-CNN realized that the region proposal phase—where the network guesses where objects might be—is often redundant. By treating existing convolutional feature maps like a physical "relief" sculpture and analyzing their intensity gradients, they can identify object boundaries almost instantaneously without a separate, heavy proposal network.
Bottom-Up Efficiency
Traditional detection is "top-down": the model looks for a bounding box and then tries to find details inside it. For specialized tasks like face detection, this is often inefficient. Yü Liu proposed KPNet, which flips this logic. By detecting facial keypoints first (bottom-up) and using them to infer the bounding box, the model achieves state-of-the-art results at a staggering 1000 FPS.
Similarly, in the realm of infrared sensing, Sylvie Le Hégarat-Mascle introduced AA-YOLO, which treats tiny targets as statistical anomalies rather than objects to be classified. This "frugal" approach allows drones to detect small heat signatures using significantly less data and power than standard segmentation models.
Distillation and Pruning
If a large "teacher" model already knows how to detect objects, can we transfer that "intuition" to a smaller "student"? This is the core of Knowledge Distillation (KD).
- Spatial Distillation: Linfeng Zhang developed PointDistiller for 3D detection, which forces a compact model to mimic the local geometric structures learned by a larger teacher.
- Temporal Distillation: Mohammad Farhadi introduced TKD, which performs this transfer online. By selecting key video frames for on-the-fly retraining, a lightweight student can maintain the accuracy of a massive "oracle" model while running at 220 FPS.
When distillation isn't enough, researchers turn to pruning—cutting away the "dead wood" of a network. Saksham Aggarwal and colleagues found that Transformers are surprisingly resilient to this process, often outperforming CNNs when both are stripped down to their bare essentials for edge deployment.
Hardware-Software Co-Design
The final frontier of this research is moving beyond the algorithm to the silicon itself. Fanrong Li and Jian Cheng demonstrated a system-level solution using FPGAs, where they combined 3-bit quantization with a custom "column-prior" tiling strategy to achieve high-speed detection at a power envelope of just 6.9W.
For the most extreme constraints, researchers are even moving away from traditional artificial neural networks. Xuemei Chen and Hangchi Shen developed SpikeSMOKE, which uses Spiking Neural Networks (SNNs) to mimic biological energy efficiency. Their architecture reduces energy consumption by 72%, making sophisticated 3D perception possible on hardware that would otherwise overheat or run out of battery in minutes.