BlankSkip: Early-exit Object Detection onboard Nano-drones

Carlo Marra; Beatrice Alessandra Motetti; Alessio Burrello; Enrico Macii; Massimo Poncino; Daniele Jahier Pagliari

cs.CVMar 30, 2026

BlankSkip: Early-exit Object Detection onboard Nano-drones

Authors: Carlo Marra, Beatrice Alessandra Motetti, Alessio Burrello, Enrico Macii, Massimo Poncino, Daniele Jahier Pagliari

Organizations: †Politecnico di Torino, Turin, Italy

Abstract

Deploying tiny computer vision Deep Neural Networks (DNNs) on-board nano-sized drones is key for achieving autonomy, but is complicated by the extremely tight constraints of their computational platforms (approximately 10 MiB memory, 1 W power budget). Early-exit adaptive DNNs that dial down the computational effort for "easy-to-process" input frames represent a promising way to reduce the average inference latency. However, while this approach is extensively studied for classification, its application to dense tasks like object detection (OD) is not straightforward. In this paper, we propose BlankSkip, an adaptive network for on-device OD that leverages a simple auxiliary classification task for early exit, i.e., identifying frames with no objects of interest. With experiments using a real-world nano-drone platform, the Bitcraze Crazyflie 2.1, we achieve up to 24% average throughput improvement with a limited 0.015 mean Average Precision (mAP) drop compared to a static MobileNet-SSD detector, on a state-of-the-art nano-drones OD dataset.

Figures & tables

Figure 1: Overview of the proposed BlankSkip architecture. The backbone processes the input frame through layers $1$ to $\ell$ . Then, the EE branch predicts the “emptyness” probability $P_{i}(1)$ . If the scene is classified as empty , inference stops early (green path), avoiding the remaining backbone layers and the SSD detection head. Otherwise, the full pipeline runs to produce bounding box predictions (red path).

Model	$\tau$	mAP	mAP ${}_{\text{no-EE}}$	Acc [%]	Skip [%]	MAC ${}_{\text{avg}}$ [M]	MAC ${}_{\text{EE}}$ [M]
Static 0.5 $\times$	–	0.412	–	–	–	193	–
Static 0.75 $\times$	–	0.479	–	–	–	358	–
Static 1.0 $\times$	–	0.591	–	–	–	534	–
EE , Stage 1 ( $\ell=1$ )	0.541	0.543	0.603	85.3	34.2	708	375
EE , Stage 2 ( $\ell=2$ )	0.717	0.533	0.587	89.6	39.2	456	178
EE , Stage 3 ( $\ell=4$ )	0.742	0.555	0.596	89.6	38.7	406	165

Table 1: Effect of EE branch positioning on Himax EE (float32).

Figure 2: Accuracy-efficiency trade-off for with threshold

\tau

sweep (0.5-0.99) on the Himax EE dataset. Stars mark optimal thresholds according to Eq. 7 . Dashed line: static SSD at different widths. Stage 1 is omitted due to prohibitively high exit costs from large feature maps (120

\times

160).

Model	mAP	Operations	Efficiency	Latency	Throughput	Model Size
		(M MAC s)	( MAC /cycle)	(ms)	(FPS)	(MB)
Static MbV2-SSD 0.5 $\times$	0.415	193	4.14	285.3	3.50	1.16
Static MbV2-SSD 0.75 $\times$	0.498	358	4.22	523.6	1.91	2.53
Static MbV2-SSD 1.0 $\times$	0.555	534	4.96	666.7	1.50	4.41
EE , Stage 4 ( $\ell=9$ )	0.540	414	4.87	537.1	1.86	4.49

Table 2: Comparison of static MobileNetV2-SSD configurations and BlankSkip on the Himax EE dataset. Int8 quantization is applied to all models. Deployment metrics are measured on GAP8.

Model	mAP	Acc	Skip	MAC ${}_{\text{avg}}$	MAC ${}_{\text{EE}}$
Static 1.0 $\times$	0.229	-	-	1807	-
EE , Stage 4 ( $\ell=10$ )	0.204	82.3	26.7	1588	855
Acc: EE branch accuracy, Skip: early-exited images [Unit = %]; MAC ${}_{\text{avg}}$ : dataset average. MAC ${}_{\text{EE}}$ : early-exit path cost (backbone + EE branch) [Unit = M $(10^{6})$ ]

Table 3: Cityscapes results (float32).

Efficient Object Detection

Efficient object detection reduces the computation, latency, memory, and energy needed to find and locate objects in images and video.

See all 95 papers

Improving Autonomous Nano-drones Performance via Automated End-to-End Optimization and Deployment of DNNs
Jul 14, 2026Vlad Niculescu, Lorenzo Lamberti, Francesco Conti +2Efficient Neural Network Inference Aerial Robotics
Integrated Systems Laboratory - ETH Zürich, Switzerland · Department of Electrical, Electronic and Information Engineering - University of Bologna, Italy · Dalle Molle Institute for Artificial Intelligence - USI-SUPSI, Switzerland
DroneScan-YOLO: Redundancy-Aware Lightweight Detection for Tiny Objects in UAV Imagery
Apr 14, 2026Yann V. BellecSmall Object Detection Aerial Object Detection
SparseVoxelDet: Fully Sparse Voxel Networks for Efficient Event-Based Drone Detection
Mar 23, 2026Mohamad Yazan Sadoun, Sarah Sharif, Yaser Mike BanadEvent-Based Object Detection Efficient Object Detection
Inquire Lab, University of Oklahoma, Norman, OK, USA.

Jul 14, 2026J/K move · Enter open · S save

Improving Autonomous Nano-drones Performance via Automated End-to-End Optimization and Deployment of DNNs

Vlad Niculescu, Lorenzo Lamberti, Francesco Conti, Luca Benini+1

Integrated Systems Laboratory - ETH Zürich, Switzerland · Department of Electrical, Electronic and Information Engineering - University of Bologna, Italy · Dalle Molle Institute for Artificial Intelligence - USI-SUPSI, Switzerland

Efficient Neural Network Inference Aerial Robotics UAV Navigation Tiny ML Vision-Based Robot Control

Open paper AskPDF arXiv

The evolution of energy-efficient ultra-low-power (ULP) parallel processors and the diffusion of convolutional neural networks (CNNs) are fueling the advent of autonomous driving nano-sized unmanned aerial vehicles (UAVs). These sub-10 cm robotic platforms are envisioned as next-generation ubiquitous smart-sensors and unobtrusive robotic-helpers. However, the limited computational/memory resources available aboard nano-UAVs introduce the challenge of minimizing and optimizing vision-based CNNs -- which to date require error-prone, labor-intensive iterative development flows. This work explores methodologies and software tools to streamline and automate all the deployment of vision-based CNN navigation on a ULP multicore system-on-chip acting as a mission computer on a Crazyflie 2.1 nano-UAV. We focus on the deployment of PULP-Dronet, a state-of-the-art CNN for autonomous navigation of nano-UAVs, from the initial training to the final closed-loop evaluation. Compared to the original hand-crafted CNN, our results show a 2x reduction of memory footprint and a speedup of 1.6x in inference time while guaranteeing the same prediction accuracy and significantly improving the behavior in the field, achieving: i) obstacle avoidance with a peak braking-speed of 1.65 m/s and improving the speed/braking-space ratio of the baseline, ii) free flight in a familiar environment up to 1.96 m/s (0.5 m/s for the baseline), and iii) lane following on a path featuring a 90 deg turn -- all while using for computation less than 1.6% of the drone's power budget. To foster new applications and future research, we open-source all the software design in a ready-to-run project compatible with the Crazyflie 2.1

Abstract

Efficient Object Detection

Explore similar work

Improving Autonomous Nano-drones Performance via Automated End-to-End Optimization and Deployment of DNNs

DroneScan-YOLO: Redundancy-Aware Lightweight Detection for Tiny Objects in UAV Imagery

SparseVoxelDet: Fully Sparse Voxel Networks for Efficient Event-Based Drone Detection