Monocular 3D Object Localization Using 2D Estimates for Industrial Robot Vision System

Thanh Nguyen Canh; Du Trinh Ngoc; Xiem HoangVan

doi:10.14313/jamris-2025-025

.blurhash-client-img { display: none !important; }

Monocular 3D Object Localization Using 2D Estimates for Industrial Robot Vision System

Journal of Automation, Mobile Robotics and Intelligent Systems

Volume 19 (2025): Issue 3 (September 2025)

By: Thanh Nguyen Canh, Du Trinh Ngoc and Xiem HoangVan

Open Access

|Sep 2025

Figures & Tables

Overview of the industrial robot vision system

Illustration of industrial robot vision system: the green point is the initialized estimate center point, and the red point is the actual center point

A block diagram of our proposed calibration method. The translation vector between the initialized estimate center point (green point), and the calibration center point (red point) is calculated based on deep learning, and our novel calibration method

The progress of the calculation of the object position in the real-world coordinate

The progress of object segmentation and edge extraction

Illustration of the estimated translation vector

Table 1.

Experiment setup details

Parameter Spec	Spec
Process	Intel Xeon Processor with two cores @ 2.3 GHz
GPU	NVIDIA Tesla T4
RAM	13 GB
OS	Ubuntu 20.04 LTS

Table 2.

Performance comparison of various object detection models

Algorithm	mAP	Pr	Rc	MS
RTMDet [37]	96.9%	94.5%	93.1%	52.3
MobileNet [35]	94.8%	93.8%	93.4%	4.6
Fast R-CNN [34]	97.0%	93.4%	94.1%	12.9
Yolov3 [33]	96.3%	95.8%	95.7%	8.7
Yolov4 [36]	96.8%	96.6%	95.4%	60.0
Yolov7 [38]	97.1%	95.7%	93.1%	37.2
Yolov8 [39]	97.8%	95.5%	94.4%	11.1
Our	98.7%	98.6%	97.0%	7.0

Visualized examples of experimental results: figure (b): the orange point is the Yolo center, figure (d): dark red is the upper part center. The vector created by the blue points is a translation vector; the light blue point is the correction center

Table 3.

Experimental results evaluate the position error of our algorithm (mm)

Fold	Sample	Traditional Method			Regression Method [30]			Proposed Method
Fold	Sample	Δx	Δy	Err	Δx	Δy	Err	Δx	Δy	Err	Δx	Δy	Err
1	1	9.69	5.51	11.15	1.34	1.32	1.88	0.38	1.14	1.20	0.95	0.76	1.22
	2	8.36	8.74	12.09	1.92	2.50	3.15	3.04	1.33	3.32	1.71	1.33	2.17
	3	5.13	8.93	10.30	1.23	1.52	1.96	1.14	0.95	1.48	1.14	0.95	1.48
	4	10.07	8.36	13.09	1.79	1.97	2.66	1.14	1.33	1.75	1.52	1.33	2.02
	5	8.55	10.26	13.36	0.75	1.12	1.35	0.76	0.19	0.78	0.38	0.57	0.69
	6	9.31	9.31	13.17	0.96	1.41	1.71	0.19	1.52	1.53	0.57	0.76	0.95
2	1	4.18	5.89	7.22	1.45	3.07	3.40	1.33	2.09	2.48	0.95	1.14	1.48
	2	10.07	13.11	16.53	3.12	3.41	4.62	2.47	1.71	3.00	2.47	1.33	2.81
	3	3.42	7.03	7.82	1.21	2.78	3.03	1.52	1.90	2.43	0.76	0.57	0.95
	4	10.07	8.74	13.33	2.13	1.51	2.61	0.95	0.19	0.97	1.90	0.95	2.12
	5	11.02	12.16	16.41	2.94	2.67	3.97	2.28	1.71	2.85	2.47	0.57	2.53
	6	8.93	11.02	14.18	0.43	1.34	1.41	1.33	0.38	1.38	0.19	0.19	0.27
3	1	9.12	7.60	11.87	1.39	1.42	1.99	0.95	0.19	0.97	1.14	0.38	1.20
	2	4.18	13.87	14.49	2.54	3.36	4.21	3.23	0.19	3.24	2.28	0.57	2.35
	3	9.88	4.18	10.73	0.76	1.12	1.35	0.57	0.76	0.95	0.57	0.76	0.95
	4	7.60	8.93	11.73	1.89	2.17	2.88	1.71	0.95	1.96	1.71	0.95	1.96
	5	10.45	4.94	11.56	0.88	2.34	2.50	1.71	1.52	2.29	0.57	0.76	0.95
	6	6.08	7.60	9.73	0.36	0.57	0.67	0.19	0.57	0.60	0.19	0.57	0.60
4	1	13.87	10.26	17.25	1.37	2.84	3.15	0.76	2.47	2.58	0.95	2.09	2.30
	2	11.97	7.98	14.39	2.84	1.45	3.19	2.28	1.33	2.64	2.28	1.33	2.64
	3	5.32	4.56	7.01	0.35	0.81	0.88	0.19	0.38	0.42	0.19	0.57	0.60
	4	5.51	16.34	17.24	0.32	1.32	1.36	0.76	1.14	1.37	0.19	1.71	1.72
	5	10.26	8.36	13.23	0.92	0.92	1.30	2.28	0.76	2.40	0.57	1.14	1.27
	6	6.27	11.40	13.01	1.47	1.63	2.19	1.90	2.85	3.43	1.33	1.33	1.88
Average		8.30	8.96	12.54	1.43	1.86	2.34	1.38	1.15	1.92	1.12	0.94	1.55

Table 4.

Performance comparison of various object segmentation models

Algorithm	mAP	Pr	Rc	MS (MB)
Yolov5 [40]	98.7%	97.1%	96.2%	7.4
RCNN [41]	97.8%	98.1%	96.4%	16.8
Yolov7 [38]	99.0%	99.0%	97.8%	37.9
Yolov8 [39]	99.2%	98.7%	97.4%	11.8
Our	99.8%	99.1%	97.9%	28.9

Table 5.

Processing time of our proposed method (milliseconds)

Phase	Processing Time
Object Detection	15 ± 2
Object Segmentation	40 ± 5
Object Calibration	300 ± 10

References

Authors

Metrics

Articles in this issue

DOI: https://doi.org/10.14313/jamris-2025-025 | Journal eISSN: 2080-2145 | Journal ISSN: 1897-8649

Journal RSS Feed

Language: English

Page range: 53 - 65

Submitted on: Jun 13, 2024

Accepted on: Sep 13, 2024

Published on: Sep 10, 2025

Published by: Łukasiewicz Research Network – Industrial Research Institute for Automation and Measurements PIAP

In partnership with: Paradigm Publishing Services

Publication frequency: 4 issues per year

Keywords:

Camera Calibration,

Object Localization,

Machine (Robot) Vision System,

Industrial Robotics

Related subjects:

Computer sciences,

Artificial intelligence,

Engineering,

Electrical engineering,

Control engineering, metrology and testing,

Mechanical engineering,

Fundamentals of mechanical engineering

© 2025 Thanh Nguyen Canh, Du Trinh Ngoc, Xiem HoangVan, published by Łukasiewicz Research Network – Industrial Research Institute for Automation and Measurements PIAP
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Volume 19 (2025): Issue 3 (September 2025)

Monocular 3D Object Localization Using 2D Estimates for Industrial Robot Vision System

Figures & Tables

Figure 1.

Figure 2.

Figure 3.

Figure 4.

Figure 5.

Figure 6.

Table 1.

Table 2.

Figure 7.

Table 3.

Table 4.

Table 5.

Paradigm

My account