Lightweight Network-Based End-to-End Pose Estimation for Noncooperative Targets

Fig. 2. Comparison of normal convolution and depthwise separable convolution. (a) Ordinary convolution; (b) depth-wise convolution; (c) point-wise convolution

Download full size

Fig. 3. Efficient channel attention structure

Download full size

Fig. 4. Basic building block of the network

Download full size

Fig. 5. Comparison of nonlinear activation functions

Download full size

Fig. 6. Examples of URSO dataset

Download full size

Fig. 7. Pose loss plot

Download full size

Fig. 8. Visual display of location estimation

Download full size

Fig. 9. Visualization display of orientation estimation. (a) Visualization of orientation ground truth; (b) visualization of predicted orientation value; (c) visualization of Euler angle estimation

Download full size

Table 1. Experimental results of fine-tuning $Δ$
View table
Table 1. Experimental results of fine-tuning $Δ$
$Δ$ $Μ$ Average orientation error /（°）
Train Test
3 16 7.6 26.0
6 16 5.5 5.8
9 16 9.3 12.4

Table 2. Influence of the number of intervals dividing the angular dimension
View table
Table 2. Influence of the number of intervals dividing the angular dimension
$M$ Network parameters /10⁶ Average orientation error /（°）
8 6.1 10.01
12 7.6 6.75
16 10.6 5.78
24 23.1 5.65

Table 3. Comparison of direct regression orientation and soft assignment coding
View table
Table 3. Comparison of direct regression orientation and soft assignment coding
Method Network parameters /10⁶ Average location error /m Average orientation error /（°）
Direct regression 5.4 0.57 46.1
Soft assignment coding 10.6 0.54 5.78

Table 4. Comparison of experiment 1
View table
Table 4. Comparison of experiment 1
Method Network parameters /10⁶ Average location error /m Average orientation error /（°） Inference time per image /ms
Method in Ref. ［25］ 45.5 1.19 13.70 45
LSPENet 10.6 0.54 5.78 39

Table 5. Comparison of experiment 2
View table
Table 5. Comparison of experiment 2
Method Network parameters/10⁶ Average location error/m Average orientation error/（°） Inference time per image/ms
Method in Ref. ［24］ 500 0.44 4.0 47
LSPENet 10.6 0.54 5.78 39

Table 6. Performance comparison of different network models
View table
Table 6. Performance comparison of different network models
Method Average location error/m Average orientation error/（°）
Deep-6DPose 1.3 13.6
Method in Ref. ［27］ 0.85 9.7
LSPENet 0.54 5.78

Table 7. Results of ablation experiments
View table
Table 7. Results of ablation experiments
ECA Average location error /m Average orientation error /（°）
0.76 8.90
√ 0.54 5.78

Table 8. Pose error on different datasets
View table
Table 8. Pose error on different datasets
Dataset Average location error /m Average orientation error /（°）
Soyuz_easy 0.54 5.78
Soyuz_hard 0.72 6.90
Dragon_hard 0.95 10.40
SPEED 0.36 5.10

Tools

Get Citation

Copy Citation Text

Jiahui Liu, Yonghe Zhang, Wenxiu Zhang. Lightweight Network-Based End-to-End Pose Estimation for Noncooperative Targets[J]. Laser & Optoelectronics Progress, 2024, 61(14): 1437010

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites