6D Pose Detection Method Based on Cross-Attention Weighting Mechanism

Yu Ye; Jing Zhang; Aimin Wang; Heng Liu; Mingju Chen

doi:10.3788/LOP250443

Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1615004(2025)

6D Pose Detection Method Based on Cross-Attention Weighting Mechanism

Yu Ye¹, Jing Zhang^1、*, Aimin Wang¹, Heng Liu¹, and Mingju Chen²

Author Affiliations

¹Southwest University of Science and Technology, School of Information Engineering, Mianyang 621010, Sichuan , China

²Artificial Intelligence Key Laboratory of Sichuan Province, Sichuan University of Science & Engineering, Yibin 643002, Sichuan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. The network architecture of the proposed pose estimation system

Download full size

Fig. 2. Architecture of the CA mechanism

Download full size

Fig. 3. Structure of main pose estimation network

Download full size

Fig. 4. Feature optimization model

Download full size

Fig. 5. Comparison of pose estimation results. (a) lamp; (b) duck; (c) driller; (d) cam; (e) cat

Download full size

Fig. 6. Comparison of validation errors during training of pose estimation networks

Download full size

Fig. 7. Accuracy threshold curve

Download full size

Fig. 8. Layout of the experimental platform. (a) Overall layout; (b) physical drawing of slide table

Download full size

Fig. 9. Pose estimation experiments. (a) Moving the measurement slider; (b) camera perspective

Download full size

Fig. 10. Pose detection in different cases. (a) Simple background; (b) complex background; (c) overlapping colors

Download full size

Fig. 11. Different background detection effects. (a) White background; (b) suit background; (c) red background

Download full size

Table 1. Comparison of ADD(-S) on the LineMOD dataset

View table

Table 1. Comparison of ADD(-S) on the LineMOD dataset

Object	DenseFusion	Method of Ref. ［16］	ORNet	TexPose	QaQ	TMFNet	Proposed method
mean	94.3	92.87	95.61	91.7	95.3	96.7	96.8
ape	92.3	85.03	81.52	80.9	90.3	93.1	95.8
benchvise	93.2	95.54	100.00	99	94.3	94.0	97.5
cam	94.4	91.27	96.86	94.8	96.8	96.5	97.9
can	93.1	95.18	98.72	99.7	95.6	98.1	94.0
cat	96.5	93.61	94.71	92.6	95.8	98.0	97.9
driller	87.0	82.56	99.01	97.4	90.0	97.0	94.5
duck	92.3	88.08	85.35	83.4	92.1	93.4	97.7
eggbox	99.80	99.90	100.00	94.9	100	99.9	99.8
glue	100.0	99.61	99.42	93.4	100	99.6	99.9
hole pucher	92.1	92.58	90.39	79.3	92.8	95.0	92.2
iron	97.0	95.91	100.00	99.8	98.1	96.9	98.2
lamp	95.3	94.43	99.42	98.3	96.9	97.9	97.8
phone	92.8	93.56	97.64	78.9	96.5	97.3	95.9

Table 2. Comparison of AUC on the YCB-Video dataset

View table

Table 2. Comparison of AUC on the YCB-Video dataset

Object	DenseFusion	D-SNet	Method of Ref. ［16］	Uni6D	Proposed method
mean	93.1	92.2	91.8	94.5	95.0
master chef can	96.4	93.3	93.9	94.8	97.4
cracker box	95.5	96.0	92.9	91.3	96.9
sugar box	97.5	97.4	95.4	95.9	97.7
tomato soup can	94.6	93.7	93.3	94.1	96.9
mustard bottle	97.2	96.1	95.4	95.2	95.8
tuna fish can	96.6	95.8	94.9	94.7	97.2
pudding box	96.5	96.2	94.0	94.3	96.3
gelatin box	98.1	95.6	97.6	97.1	98.4
potted meat can	91.3	89.3	90.6	92.8	95.3
banana	96.6	96.9	91.7	96.0	97.4
pitcher base	97.1	96.3	93.1	96.3	97.3
bleach cleanser	95.8	94.0	93.4	94.7	95.9
bowl	88.2	85.1	92.9	94.3	93.5
mug	97.1	97.1	96.1	96.5	97.9
power drill	96.0	94.5	93.3	93.8	96.3
wood block	89.7	90.9	87.6	94.4	93.6
scissors	95.2	93.8	95.7	86.9	94.8
large marker	97.5	96.8	95.7	96.4	97.0
large clamp	72.9	72.0	75.4	94.1	82.6
extra large clamp	69.8	73.2	73.0	93.9	81.1
foam brick	92.5	91.8	94.2	96.0	96.2

Table 3. Comparative analysis of model ablation experiments
View table
Table 3. Comparative analysis of model ablation experiments
RGB-D stage 1 stage 2 stage 3 baseline model ADD（-S） /%
√ √ 88.8
√ √ √ 93.7
√ √ √ √ 94.9
√ √ √ √ √ 96.8

Table 4. Detection accuracy under different color backgrounds
View table
Table 4. Detection accuracy under different color backgrounds
Color background Detection accuracy
Blue background 96.6
White background 96.3
Suit background 95.9
Red background 96.1

Table 5. Comparison of centroid position and bias estimation results of different methods

View table

Table 5. Comparison of centroid position and bias estimation results of different methods

Target object	Setting position	Position estimation		Devaton of cemtoid
Target object	Setting position	DenseFusion	Proposed method	DenseFusion	Proposed method
stapler	（2.00，-9.67，6.30）	（1.11，-9.05，4.35）	（1.86，-9.55，4.71）	2.23	1.60
stapler	（1.00，-6.67，6.28）	（1.82，-7.93，5.12）	（1.82，-7.23，5.50）	1.90	1.27
globule	（2.00，8.67，3.11）	（2.73，7.93，1.68）	（2.20，8.03，1.70）	1.75	1.56
globule	（-1.00，-7.67，3.11）	（-1.79，-6.01，1.81）	（-1.62，-6.04，2.21）	2.25	1.99
brick	（2.00，-9.67，3.65）	（2.84，-7.82，2.50）	（2.54，-8.35，2.80）	2.33	1.66
brick	（1.00，-6.67，3.65）	（1.96，-5.15，2.37）	（1.66，-5.45，2.78）	2.21	1.64
mean				2.11	1.62

Table 6. Rotation angle deviation measurement

View table

Table 6. Rotation angle deviation measurement

Angle /（°）	X /mm	Y /mm	Z /mm	A /（°）	B /（°）	C /（°）	Angular error /（°）
0	12.0	6.1	5.2	1.19	1.19	359.15	0.85
90	11.8	6.0	5.0	1.92	1.19	90.27	0.27
180	11.5	5.7	4.7	1.66	0.80	181.39	1.39
270	11.6	5.8	4.8	1.38	1.90	269.48	0.52

Tools

Get Citation

Copy Citation Text

Yu Ye, Jing Zhang, Aimin Wang, Heng Liu, Mingju Chen. 6D Pose Detection Method Based on Cross-Attention Weighting Mechanism[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1615004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites