Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0415002(2025)
Adaptive Multimodal-Feature Fusion for 6D Object Position Estimation
Figures & Tables(9)
Fig. 5. Comparison of pose estimation results. (a) ape; (b) eggbox; (c) cat; (d) camera
Fig. 6. Experimental results of surface normal generation from depth information. (a) Original image; (b) normal image; (c) normal image of the target object
Table 1. Test results on LineMOD dataset
View tableTable 1. Test results on LineMOD dataset
Object RGB based method RGB-D based method PoseCNN PVNet HybridPose TexPose DenseFusion MaskedFusion Uni6D QaQ Proposed method mean 88.6 86.3 94.5 91.7 94.3 97.3 97.0 95.3 97.4 ape 77.0 43.6 77.6 80.9 92.3 92.2 93.7 90.3 96.1 bench yi. 97.5 99.9 99.6 99.0 93.2 98.4 99.8 94.3 97.6 camera 93.5 86.9 95.9 94.8 94.4 98.0 96.0 96.8 97.3 can 96.5 95.5 93.6 99.7 93.1 97.4 99.0 95.6 96.9 cat 82.1 79.3 93.5 92.6 96.5 97.8 98.1 95.8 97.7 driller 95.0 96.4 97.2 97.4 87.0 95.6 99.1 90.0 94.5 duck 77.7 52.6 87.0 83.4 92.3 94.0 99.0 92.1 95.8 eggbox 97.1 99.2 99.6 94.9 99.8 99.6 100.0 100.0 99.8 glue 99.4 95.7 98.7 93.4 100.0 100.0 99.2 100.0 99.9 hole p. 52.8 82.0 92.5 79.3 92.1 97.3 90.2 92.8 96.5 iron 98.3 98.9 98.1 99.8 97.0 97.1 99.5 98.1 99.0 lamp 97.5 99.3 96.9 98.3 95.3 99.0 99.4 96.9 97.8 phone 87.7 92.4 98.3 78.9 92.8 98.8 97.4 96.5 97.4
Table 2. Test results on YCB-Video dataset
View tableTable 2. Test results on YCB-Video dataset
Object DenseFusion Dual-Stream PR-GCN Uni6D Proposed method AUC <2 cm /% AUC <2 cm /% AUC <2 cm /% AUC <2 cm /% AUC <2 cm /% mean 91.6 96.2 92.2 96.5 95.0 97.7 94.5 97.5 95.8 98.3 master chef can 96.0 100.0 93.3 99.8 95.6 99.4 94.8 98.7 96.3 100.0 cracker box 93.1 99.2 96.0 99.4 97.0 100.0 91.3 98.4 97.3 100.0 sugar box 96.8 99.6 97.4 99.9 97.8 98.1 95.9 99.1 98.1 99.8 tomato soup can 93.3 95.7 93.7 96.5 94.6 97.2 94.1 96.4 95.8 97.4 mustard bottle 97.0 99.8 96.1 100.0 98.0 99.6 95.2 99.0 98.6 100.0 tuna fish can 95.9 100.0 95.8 99.5 96.4 99.8 94.7 98.8 97.2 100.0 pudding box 94.8 99.4 96.2 99.7 97.8 99.2 94.3 99.2 97.9 99.8 gelatin box 98.0 100.0 95.6 95.2 96.4 95.3 97.1 99.8 98.2 100.0 potted meat can 88.7 92.1 89.3 91.4 95.1 97.9 92.8 94.0 95.4 94.9 banana 94.3 98.9 96.9 100.0 97.1 99.9 96.0 99.4 98.0 100.0 pitcher base 96.9 99.3 96.3 98.9 97.6 100.0 96.3 99.5 98.3 100.0 bleach cleanser 94.2 98.9 94.0 99.9 96.3 99.1 94.7 97.6 96.8 99.6 bowl 85.1 97.6 85.1 99.3 90.5 96.8 94.3 96.9 94.1 100.0 mug 97.2 99.9 97.1 98.5 97.5 99.3 96.5 99.7 97.9 99.8 power drill 94.0 97.2 94.5 98.1 97.0 99.0 93.8 95.3 97.4 98.5 wood block 87.8 93.0 90.9 97.7 94.8 97.7 94.4 97.2 95.3 97.9 scissors 93.7 98.7 93.8 99.0 94.9 98.5 86.9 84.6 94.2 100.0 large marker 97.2 100.0 96.8 99.8 96.8 99.4 96.4 99.8 97.0 100.0 large clamp 71.6 78.1 72.0 77.2 87.2 92.6 94.1 98.3 88.7 90.1 extra large clamp 66.0 74.7 73.2 76.9 79.1 83.3 93.9 97.5 83.4 87.4 foam brick 92.2 97.5 91.8 98.9 96.0 98.9 96.0 99.2 96.5 100.0
Table 3. Comparative analysis of model ablation experiments
View tableTable 3. Comparative analysis of model ablation experiments
RGB-D Improve feature extraction Normal enhancement module Feature fusion module Baseline model ADD /% √ √ 88.8 √ √ √ 93.4 √ √ √ √ 95.9 √ √ √ √ √ 97.4
Tools
Get Citation
Copy Citation Text
Chuanfang Zang, Jianwu Dang, Jiu Yong. Adaptive Multimodal-Feature Fusion for 6D Object Position Estimation[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0415002
Paper Information
Category: Machine Vision
Received: May. 8, 2024
Accepted: Jun. 17, 2024
Published Online: Feb. 10, 2025
The Author Email:
CSTR:32186.14.LOP241238
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20