# CucumberVision study measures cucumber sizing on Jetson Orin Nano and Raspberry Pi 4

> Researchers at the University of Windsor and MBZUAI report accuracy, throughput and energy for a cucumber-length pipeline on two low-power edge boards.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: AI in Practice
- Published: 2026-11-16T14:01:15+00:00
- Updated: 2026-11-16T14:09:24.540416+00:00
- Web page: https://ai-news.ikangai.com/story/cucumbervision-study-measures-cucumber-sizing-on-jetson-orin-nano-and-raspberry/
- Seen in: Import AI (Jack Clark)
- Change: research (Research)

Researchers from the University of Windsor and the Mohamed bin Zayed University of Artificial Intelligence report that a cucumber-length measurement pipeline runs on an NVIDIA Jetson Orin Nano with the same mean absolute error as an RTX 3050 Ti workstation: 0.58 cm on both, with a paired difference of zero. The authors, Manveen Kaur, Rajmeet Singh, Saeed Mozaffari, Majid Ahmadi and Shahpour Alirezaee, describe the work in a preprint submitted on 6 November 2026.

The pipeline has three stages. A YOLO26n detector finds the fruit in an RGB-D image. A prompt-based segmentation model isolates its outline. A medial arc spline estimator then fits a curved centreline along the mask to produce a length. The authors ran 50 identical greenhouse captures of seven specimens through each setup. They used specimen-level cluster-bootstrap statistics to put a confidence interval on the error, which runs from 0.46 to 0.69 cm. The accuracy figures describe these replayed captures, not in-field grading.

The speed and energy results depend on which segmentation backend is used, and the paper separates the two effects. According to the authors, accuracy is set by the backend and reproduces across platforms to within 0.01 cm. The cost of running that backend is set by the hardware, and it varies by 214 times. On the Jetson, TensorRT FP16 compilation cut detection latency 2.57 times and produced two operating points: 72 ms per fruit at 0.84 cm of error, or 2.74 s per fruit at 0.64 cm. Energy per fruit on the Jetson was 42.2 J. Its 7 W power mode consumed 73% more than its 15 W mode.

In the paper's pairing, the Raspberry Pi 4 costs more per fruit. Running MobileSAM on the Pi cost 2.5 times as much per fruit as running SAM on the Jetson, even though the Pi drew 47% less power. The authors say the Pi ran without thermal throttling under sustained load. That matters for a scouting platform with no active cooling.

The paper also reports a failure. The authors tried whole-graph INT8 quantization on the detector's NMS-free head, the step that normally shrinks a model for faster integer arithmetic. The detector returned no detections at all. The paper attributes this to one output tensor that carries both pixel coordinates and confidence scores, so quantizing the tensor as a whole fails. The authors claim that no calibration strategy can avoid this. Excluding the head from quantization restored function at a 1.37 times speedup. The paper presents this as a structural property of this detector design, not a general failure of INT8 on edge boards.

Taken together, the paper defines what it calls an accuracy-throughput-energy envelope with three non-dominated configurations. No one of them is beaten on every measure by another. The paper also gives selection rules for choosing among them. The authors' summary is that the Jetson with native masks suits throughput, the Jetson with SAM suits accuracy, and the Raspberry Pi with native masks is the low-cost option. The paper's data come from replayed captures of seven specimens, so the configuration ranking is specific to these measurements.

## Why it matters

The study gives a grounded benchmark for running computer vision on low-power hardware in the field, which matters for teams deploying AI outside data centres.

## Original source

- [(2611.09490) CucumberVision at the Edge: Accuracy, Throughput, and Energy of RGB-D Cucumber Sizing on Jetson Orin Nano and Raspberry Pi 4](https://arxiv.org/abs/2611.09490)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "CucumberVision study measures cucumber sizing on Jetson Orin Nano and Raspberry Pi 4", 2026-11-16, https://ai-news.ikangai.com/story/cucumbervision-study-measures-cucumber-sizing-on-jetson-orin-nano-and-raspberry/
