Input-Adaptive Dynamic Neural Network for Efficient Object Detection Toward Resource-Constrained Deployment

The deployment of object detection models on resource-constrained edge devices remains a substantial challenge, primarily because conventional static networks expend the same worst-case computational cost on every input, regardless of intrinsic difficulty. This paper proposes an input-adaptive dynamic neural network architecture for object detection in embedded environments. The present study investigates two orthogonal axes of input-adaptive inference for embedded object detection: The system demonstrates depth adaptivity through the implementation of Early Exit, and width adaptivity via group-wise Adaptive Routing. The proposed framework is constructed on a frozen Ultralytics YOLO26s backbone and incorporates two YOLO-style early-exit heads positioned at approximately 33% and 66% of the backbone depth. Furthermore, the framework incorporates two Straight-Through Gumbel-Softmax routers, which are appended after Layers 4 and 8 with group-wise hard gating. Both axes additionally accept an explicit external control signal that allows the host system to override the input-conditional policy at inference time. The dual-mode design facilitates the functionality of the trained checkpoint as either an input-adaptive policy, in which the depth and width are determined per sample from the input distribution, or an externally controlled policy. The experimental findings demonstrate two strongly asymmetric input-adaptive policies on a frozen YOLO26s backbone. The early-exit profile reduces the compute per sample from 12.739 to 10.532 GFLOPs—a 17.32% reduction according to our in-house Conv2d/Linear MAC-based GFLOPs estimator—while preserving baseline accuracy (mAP50 = 0.1545 vs. baseline = 0.1528; ΔmAP50 = +0.0017, within evaluation noise; mAP50–95 = −0.0033). Evaluating the router-only profile in the same validator pipeline with a sparsity penalty of γ = 0.05 results in a 12.3% decrease in logical GFLOPs (from 12.739 to 11.172), while maintaining an accuracy level that is at or above the PEFT baseline (mAP50 = 0.2324 and mAP50–95 = 0.1040). In our small-domain PEFT setup, training the dynamic-policy modules yields per-checkpoint mAP shifts in this magnitude. Therefore, we interpret the width-axis accuracy result as preservation of the baseline rather than an improvement. Our contribution on the width axis is reducing computing power while maintaining baseline accuracy. Importantly, the router profile’s logical GFLOP savings are not currently reflected in wall-clock latency under our dense-kernel PyTorch implementation. Achieving practical speedup requires sparse-kernel deployment, such as structured-sparse kernels, TensorRT, TVM, or Triton paths. We will address this in future deployment-level work. Our results indicate that the depth axis can yield genuine end-to-end speedup today, while the width axis offers deployment-pending compute reduction.