Technical Summary: MITE-Net: SWaP-Optimized 4K Video Tiny Target Perception for Embodied Edge SAR

1. Problem Statement

Real-time perception of tiny targets in high-resolution (4K) imagery is critical for embodied Search-and-Rescue (SAR) missions, such as locating stranded individuals in maritime environments or identifying ignition sources in wildfires. However, deploying these capabilities on edge devices (e.g., UAVs) faces a fundamental bottleneck driven by strict Size, Weight, and Power (SWaP) constraints.

Existing approaches struggle to balance three competing factors:

1. Direct Downsampling: Traditional detectors resize 4K inputs to lower resolutions (e.g., 640×640) to maintain speed, but this destroys the fine-grained features of targets that may occupy only a few dozen pixels.
2. Slice-Based Inference: Methods like SAHI preserve local details by processing image patches but incur prohibitive latency and computational overhead, failing to meet real-time requirements.
3. Ultra-Lightweight Models: While models like YOLO-Nano are efficient, they often suffer from "representational collapse," lacking the capacity to simultaneously suppress background noise and extract fragile features of sub-64-pixel targets in complex scenes.

The paper posits that a decoupled approach is necessary to bridge the gap between high-frequency feature preservation and real-time inference under strict energy budgets.

2. Methodology: MITE-Net Architecture

The authors propose MITE-Net (Motion-Informed Tiny-target Edge Network), a cascaded, two-stage architecture designed to decouple spatial perception from semantic recognition.

Stage 1: Bio-Inspired, Learning-Free TTM-RPN

The first stage utilizes a Tiny Target Motion-based Region Proposal Network (TTM-RPN) built upon the vSTMD (various-velocity Small Target Motion Detector).

• Mechanism: Instead of learning features, this module uses bio-inspired spatio-temporal filtering (mimicking dragonfly vision) to detect motion cues. It operates on a downsampled space to minimize computational cost.
• Function: It generates a score map and direction map to isolate moving objects from static backgrounds. It outputs a sparse set of Regions of Interest (RoIs) based on motion saliency.
• Motion-Enhanced Cropping: The system crops not only the raw 4K image but also temporal change maps and motion response maps. These are aligned and concatenated to form a multi-modal input tensor that preserves high-resolution spatial details while incorporating polarity-aware motion cues.
• Configurable Downsampling: To save power, the motion detection can operate on downsampled frames (e.g., 1920×1080), while the subsequent cropping is performed on the native 4K resolution to ensure no microscopic detail is lost.

Stage 2: Ultra-Lightweight R-CNN-like Head

The second stage processes the cropped RoIs (resized to 16×16 pixels) using a highly specialized detection head with <0.14M parameters.

• Architecture: It employs a dilated backbone to expand the receptive field without pooling (preserving tiny target features), Coordinate Convolution (CoordConv) for spatial awareness, and decoupled heads for classification and regression.
• Loss Functions: The training strategy utilizes Normalized Wasserstein Distance (NWD) for robust bounding box regression (superior to IoU for tiny targets) and Distribution Focal Loss (DFL) to handle boundary ambiguity.
• Training Strategy: The system employs Ground Truth jittering to simulate imperfect motion proposals and Hard Negative Mining to suppress false alarms generated by the motion detector.

3. Key Contributions

The paper outlines three primary contributions:

1. SWaP-Optimized Architecture: MITE-Net successfully decouples background filtering (via bio-inspired motion detection) from semantic perception. This allows the system to run a computationally intensive, high-resolution recognition head only on sparse, motion-salient regions, achieving real-time performance under strict power constraints.
2. SAR-Tiny Datasets: The authors constructed and open-sourced two new 4K datasets to standardize evaluation:
   • SeaDroneSee-Tiny: Focuses on dynamic maritime environments with targets predominantly in the 64–256 pixel range, though the test set specifically stresses the boundary with "Extremely Tiny" (≤64 pixels) targets.
   • UAVID-Tiny: Focuses on highly cluttered urban scenes with extremely tiny targets (≤64 pixels).
3. Hardware-Level Benchmarks: The work provides rigorous benchmarks on the NVIDIA Jetson AGX Xavier, establishing critical baselines against state-of-the-art YOLO models (v11n, v8n-P2) and demonstrating the structural boundaries of current lightweight models.

4. Experimental Results

Experiments were conducted on the NVIDIA Jetson AGX Xavier using 4K maritime and urban imagery.

Performance on SeaDroneSee-Tiny (Maritime)

• Success Rate: MITE-Net achieved a 100% Search Success Rate (SSR) at 30.33 FPS, directly processing raw 4K imagery. In contrast, resized YOLO baselines dropped to 83.33% due to feature loss.
• Efficiency: The system achieved an inference power consumption of 3.19 W, resulting in an energy efficiency of 9.51 FPS/Watt. This vastly outperformed YOLO baselines, which required significantly more power for lower or comparable success rates. (Note: The total system power consumption was 21.10 W, but the efficiency metric is derived from the inference-specific power).
• Trade-off: While MITE-Net exhibited a higher False Alarm Rate (FAR) in the maritime setting (67.72%), the authors argue that in life-critical SAR, preventing missed detections is prioritized over bounding box precision.

Performance on UAVID-Tiny (Urban)

• Limitations: In highly cluttered urban scenes, MITE-Net's performance degraded significantly, achieving only 8.33% SSR (and 0% under aggressive downsampling).
• Analysis: The authors attribute this to a "compound structural limitation": the learning-free bionic front-end struggles to distinguish targets from complex urban motion noise, and the ultra-lightweight head lacks the representational capacity to classify targets in such dense, static-heavy environments.

5. Significance and Claims

The paper claims that MITE-Net delivers an efficient onboard perception paradigm that resolves the feature loss dilemma inherent in downsampling 4K imagery. By leveraging bio-inspired motion dynamics to filter backgrounds before deep learning inference, the system enables highly sensitive target perception under strict edge SWaP constraints.

The authors modestly acknowledge that while the approach is superior for dynamic, open environments (like maritime SAR), it exposes the limitations of ultra-lightweight, motion-dependent architectures in hyper-cluttered static scenes. Consequently, this work serves not only as a solution for specific SAR scenarios but also as a rigorous baseline that maps the representational boundaries of current edge AI, guiding the development of future end-to-end SAR architectures.