EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Lightweight Monocular Depth Estimation with Depth Anything for Real-Time Deployment on Resource-Constrained Platforms

Jiayu Huang, Yiduo Chen

Sensors · 11.11.2026 · 26 (22) · 7121

DOI: 10.3390/s26227121 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Sensors, online erschienen am 11.11.2026. Verfasst von Jiayu Huang und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Monocular depth estimation enables low-cost 3D perception using RGB cameras, but deployment on resource-constrained platforms requires a balance between accuracy and efficiency. Existing approaches face four challenges: computational constraints, representation degradation, inference latency, and scale uncertainty. Building on Depth Anything ViT-S/14, we prune complete Transformer blocks as the minimal unit in a nested design with shallower models as subsets of deeper ones. Through a controlled 6/8/10-block ablation, we select an 8-block model retaining multi-scale Dense Prediction Transformer (DPT) decoder features. Metric-aware fine-tuning with per-image median scale alignment improves post-pruning accuracy. Evaluation uses least-squares scale-and-shift alignment. TensorRT FP16 inference is deployed on the Jetson Orin Nano Super using Open Neural Network Exchange (ONNX) conversion. Under matched training conditions and the scale-and-shift alignment protocol, the 8-block model achieves a benchmark absolute relative error (AbsRel) of 0.1350 with 28.7% fewer parameters than the 12-block reference, yielding a reasonable accuracy-efficiency trade-off. After fine-tuning on the NYU Depth Dataset V2 (NYUv2), AbsRel is 19.1% lower than that of the official zero-shot baseline. On Jetson, TensorRT FP16 reaches 240.26 FPS at the model level (1.33 times the 12-block reference) but approximately 25 FPS in the camera end-to-end pipeline (∼40 ms latency). The prototype runs stably for 30 min with GPU temperature below 57.5 ∘C. Under controlled fixed-camera conditions, the stop-gradient (sg) variant achieves a scale-recovery mean absolute error (MAE) of 0.0674 m (PC) and 0.0760 m (Jetson) at four held-out distances within 0.5–1.5 m. Structured pruning with metric-aware fine-tuning balances accuracy and deployment efficiency while retaining pretrained block weights, providing a compression-to-deployment workflow for edge monocular depth estimation under controlled conditions.

Zitieren

APA

Huang, J., & Chen, Y. (2026). Lightweight Monocular Depth Estimation with Depth Anything for Real-Time Deployment on Resource-Constrained Platforms. Sensors, 26(22), 7121. https://doi.org/10.3390/s26227121

Vancouver

Huang J, Chen Y. Lightweight Monocular Depth Estimation with Depth Anything for Real-Time Deployment on Resource-Constrained Platforms. Sensors. 2026;26(22):7121. doi:10.3390/s26227121

BibTeX
@article{Huang2026,
  author = {Huang, Jiayu and Chen, Yiduo},
  title = {Lightweight Monocular Depth Estimation with Depth Anything for Real-Time Deployment on Resource-Constrained Platforms},
  journal = {Sensors},
  year = {2026},
  volume = {26},
  number = {22},
  pages = {7121},
  doi = {10.3390/s26227121}
}
Kostenlos starten →