CucumberVision at the Edge: Accuracy, Throughput, and Energy of RGB-D Cucumber Sizing on Jetson Orin Nano and Raspberry Pi 4
Organizations: Department of Mechanical Engineering, University of Windsor, Windsor, ON, Canada · Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Department of Electrical and Computer Engineering, University of Windsor, Windsor, ON, Canada
Abstract
Non-contact fruit sizing from RGB-D imagery reaches sub-centimetre accuracy on workstation hardware, but in-field grading requires the same pipeline to run on a low-power single-board computer, and the cost of that transition is rarely reported. We deploy a curvature-aware cucumber length-measurement pipeline (YOLO26n detection, prompt-based segmentation and a medial arc spline estimator) on an NVIDIA Jetson Orin Nano and a Raspberry Pi 4, and characterize both against an RTX 3050 Ti workstation by replaying 50 identical greenhouse captures of 7 specimens, with specimen-level cluster-bootstrap statistics. Accuracy is preserved: mean absolute error (MAE) is 0.58 cm (95% CI 0.46-0.69) on both the workstation and the Jetson at matched precision, with a paired difference of zero. TensorRT FP16 compilation cuts Jetson detection latency 2.57x, giving two operating points on one board: 72 ms per fruit at 0.84 cm, or 2.74 s at 0.64 cm. Four ablations map the design space. For prompted segmentation backends, measurement accuracy is fixed by the backend and reproduces across platforms to within 0.01 cm, while its cost is fixed by the platform and varies 214x, so the two choices are separable. Whole-graph INT8 quantization of the NMS-free detection head fails structurally, returning no detections, because one output tensor carries both pixel coordinates and confidences; we show that no calibration strategy can avoid this. Excluding the head restores function at 1.37x. Energy per fruit is 42.2 J on the Jetson, and its 7 W mode consumes 73% more than its 15 W mode. The Raspberry Pi with MobileSAM costs 2.5x more per fruit than the Jetson with SAM despite drawing 47% less power, and it runs without thermal throttling under sustained load. The results define an accuracy-throughput-energy envelope with three non-dominated configurations and give rules for selecting among them.
Figures & tables
| Workstation | Jetson Orin Nano | RPi 4 B | |
|---|---|---|---|
| Compute | RTX 3050 Ti | Ampere, 1024 cores | Cortex-A72 CPU |
| Runtime | PyTorch CUDA | PyTorch / TensorRT | ONNX Runtime |
| Precision | FP32 | FP32, FP16 | FP32, INT8 |
| Software | — | JetPack 6.0, TensorRT 8.6.2 | RPi OS 64-bit |
| Cooling | Active | Active | Passive |
| Configuration | Det. | Seg. | Per fruit | MAE | RMSE | RAM |
|---|---|---|---|---|---|---|
| (ms) | (ms) | (s) | (cm) | (cm) | (MB) | |
| Laptop, SAM ViT-B | 43 | 421 | 0.47 | 0.58 | 0.88 | 1466 |
| Laptop, MobileSAM | 46 | 70 | 0.12 | 0.76 | 1.13 | 1568 |
| Laptop, native | 46 | — | 0.05 | 0.98 | 1.29 | 2059 |
| Jetson FP32, SAM | 120 | 2653 | 2.79 | 0.58 | 0.88 | 1272 |
| Jetson TRT FP16, SAM | 46 | 2680 | 2.74 | 0.64 | 0.97 | 1344 |
| Configuration | MAE (cm) | 95% CI | Bias | 95% LoA | |
|---|---|---|---|---|---|
| Laptop, SAM ViT-B | 50 | 0.58 | 0.46–0.69 | +0.06 | to |
| Laptop, MobileSAM | 50 | 0.76 | 0.56–1.00 | to | |
| Laptop, native | 50 | 0.98 | 0.66–1.24 | to | |
| Jetson FP32, SAM | 50 | 0.58 | 0.46–0.69 | +0.06 | to |
| Jetson TRT FP16, SAM | 50 | 0.64 | 0.55–0.73 | +0.09 | to |
| Jetson TRT FP16, native | 50 | 0.84 | 0.69–0.98 | +0.57 | to |
| Configuration | Det. | Speed- | Det- | MAE | Model | RAM |
|---|---|---|---|---|---|---|
| (ms) | up | ects | (cm) | (MB) | (MB) | |
| FP32, wrapper | 3266 | yes | — | 41.7 | — | |
| FP32, direct | 1929 | yes | 0.82 | 41.7 | 682 | |
| INT8, head FP32 | 1406 | yes | 1.12 | 15.1 | 594 | |
| INT8, whole graph | 793 | no | — | 11.1 | 460 |
| Configuration | Mean draw (W) | Run time (s) | J/fruit | J/fruit, net | M5 MAE (cm) |
|---|---|---|---|---|---|
| Jetson TensorRT FP16 + SAM, 15 W | 10.14 | 183 | 42.18 | 20.18 | 0.64 |
| Jetson TensorRT FP16 + SAM, 7 W | 7.21 | 446 | 73.12 | 23.45 | 0.64 |
| RPi 4 FP32 + MobileSAM | 5.41 | 859 | 105.55 | 56.38 | 0.76 |
| RPi 4 FP32, native masks | 5.68 | 117 | 15.05 | 8.38 | 0.82 |
| RPi 4 INT8 head-FP32, native masks | 5.61 | 89 | 11.29 | 6.22 | 1.12 |
| Comparison | Det. | Box IoU | Mask IoU | Shift |
|---|---|---|---|---|
| (%) | mean (min) | mean (min) | px | |
| Jetson FP32 vs laptop, SAM | 100 | 1.000 (1.000) | 1.000 (1.000) | 0.00 |
| Jetson TRT vs Jetson FP32 | 100 | 0.990 (0.946) | 0.999 (0.992) | 0.69 |
| RPi vs laptop, MobileSAM | 100 | 0.994 (0.966) | 1.000 (0.997) | 0.54 |
| RPi vs laptop, native | 100 | 0.994 (0.966) | 0.823 (0.447) | 0.54 |
| MobileSAM vs SAM, laptop | 100 | 1.000 (1.000) | 0.957 (0.788) | 0.00 |
| Configuration | M5 MAE (cm) | Per fruit (s) | J/fruit | Non- dominated |
|---|---|---|---|---|
| Laptop, SAM ViT-B (reference) | 0.58 | 0.47 | — | ref. |
| Jetson TensorRT FP16, native masks | 0.84 | 0.07 | 0.7 | yes |
| RPi 4 INT8 head-FP32, native masks | 1.12 | 1.44 | 11.3 | no |
| RPi 4 FP32, native masks | 0.82 | 1.96 | 15.1 | yes |
| Jetson TensorRT FP16, SAM | 0.64 | 2.74 | 42.2 | yes |
| RPi 4 INT8, MobileSAM † | 0.76 | 15.8 | 87.8 | no |
Agricultural Image Analysis
Agricultural image analysis uses computer vision and multimodal learning to monitor crops, livestock, produce quality, and field conditions.
See all 103 papers