Chinese Optics Letters, Volume. 23, Issue 9, 091101(2025)
Image-free cross-species pose estimation via an ultra-low sampling rate single-pixel camera
Figures & Tables(9)
Fig. 1. Overview of plane-array-camera-based and SPC-based pose estimation models. (a) The architecture of plane-array-camera-based (image-based) pose estimation model. (b) The architecture of SPC-based (image-free) pose estimation model that performs reconstruction before pose estimation, the two-stage architecture of image-free model with implicit reconstruction, and the architecture of our proposed image-free and (implicit) reconstruction-free model.
Fig. 2. Overview of the SPCPose. (a) SPC structure schematic. (b) The architecture of our proposed image-free and (implicit) reconstruction-free cross-species pose estimation model.
Fig. 3. Visualization of SPCPose on the public dataset. (a) Visualization of SPCPose on Tri-Mouse. (b) Visualization of SPCPose on Horse10.
Fig. 4. Visualization of SPCPose on the indoor Human dataset. The numbers on the picture mean the sampling rate.
Fig. 5. Real-world experimental results. (a) Schematic of our customized SPC and captured scenes. (b) Results of the same object captured with an RGB camera perform different actions in three real-world scenes. (c) Reconstructed scenes using our single-pixel detection system and pose estimation results using an image-based approach. (d) Results of the natural-order single-pixel detection value map representation and the action of SPCPose. (e) Inverse-order single-pixel detection value map representation and results from the action of SPCPose. (f) Random-order single-pixel probe-value map representation and results of the action of SPCPose.
Fig. 6. Effect of extraction methods on SPCPose’s performance in parsing object poses at 256 sample points. (a) Visualization of SPCPose processing results with different extraction methods. The numbers on the picture mean the sampling rate. (b) Objective evaluation of SPCPose processing performance with different extraction methods.
Table 1. Performance Comparison of Image-Based and Image-Free Pose Estimation Regarding AP, PCK, AUC, EPE, Params, and GFLOPs on Tri-Mouse[6,51]
a View tableView in ArticleTable 1. Performance Comparison of Image-Based and Image-Free Pose Estimation Regarding AP, PCK, AUC, EPE, Params, and GFLOPs on Tri-Mouse[6,51]
a Method Sample rate Backbone AP@50-95 PCK@0.05 AUC EPE Params (M) GFLOPs Heatmap[ 52 ]1.000 HRNet-W32 99.9 99.8 91.4 1.91 28.5 10.2 1.000 HRNet-W48 100.0 99.8 92.2 1.67 63.6 21.0 RLE[ 48 ]1.000 ResNet50 98.4 95.8 86.4 3.43 23.7 5.4 1.000 ResNet101 99.3 97.4 87.7 3.01 42.7 10.2 1.000 ResNet152 98.9 96.4 87.5 3.08 58.3 15.1 SPCPose-M1 5.333 × 10−2 ViT-S 99.9 95.9 91.4 1.90 24.3 5.9 1.333 × 10−2 ViT-S 96.5 99.9 88.5 3.56 24.3 5.9 3.333 × 10−3 ViT-S 53.0 54.8 54.5 25.39 24.3 5.9 8.333 × 10−4 ViT-S 78.4 79.7 77.6 14.59 24.3 5.9 SPCPose-M2 1.333 × 10−2 ViT-S 100.0 100.0 93.9 1.13 24.3 5.9 3.333 × 10−3 ViT-S 100.0 100.0 92.8 1.49 24.3 5.9 8.333 × 10−4 ViT-S 100.0 100.0 93.4 1.30 24.3 5.9 SPCPose-M3 1.333 × 10−2 ViT-S 100.0 99.7 91.5 1.87 24.3 5.9 3.333 × 10−3 ViT-S 100.0 99.8 91.3 1.93 24.3 5.9 8.333 × 10−4 ViT-S 100.0 100.0 93.4 1.28 24.3 5.9
Table 2. Performance Comparison of Image-Based and Image-Free Pose Estimation Regarding AP, PCK, AUC, EPE, Params, and GFLOPs on Horse10[51]
a View tableView in ArticleTable 2. Performance Comparison of Image-Based and Image-Free Pose Estimation Regarding AP, PCK, AUC, EPE, Params, and GFLOPs on Horse10[51]
a Method Sample rate Backbone AP@50-95 PCK@0.05 AUC EPE Params (M) GFLOPs Heatmap[ 52 ]1.000 HRNet-W32 98.3 99.9 93.6 1.14 28.5 10.2 1.000 HRNet-W48 98.3 99.9 93.7 1.10 63.6 21.0 RLE[ 48 ]1.000 ResNet50 95.1 99.6 93.8 1.44 23.7 5.4 1.000 ResNet101 96.1 99.7 93.9 1.36 42.7 10.2 1.000 ResNet152 96.5 99.7 94.0 1.30 58.3 15.1 SPCPose-M1 3.512 × 10−1 ViT-S 87.6 94.0 88.1 2.96 24.3 5.9 8.779 × 10−2 ViT-S 76.0 86.6 83.0 4.85 24.3 5.9 2.195 × 10−2 ViT-S 67.4 81.5 80.2 5.83 24.3 5.9 5.487 × 10−3 ViT-S 72.6 83.2 81.0 5.65 24.3 5.9 SPCPose-M2 8.779 × 10−2 ViT-S 92.7 97.7 91.3 1.87 24.3 5.9 2.195 × 10−2 ViT-S 89.4 94.9 88.5 2.83 24.3 5.9 5.487 × 10−3 ViT-S 83.3 90.3 85.0 4.10 24.3 5.9 SPCPose-M3 8.779 × 10−2 ViT-S 94.7 98.1 93.4 1.52 24.3 5.9 2.195 × 10−2 ViT-S 91.1 96.5 90.3 2.20 24.3 5.9 5.487 × 10−3 ViT-S 89.6 95.8 89.6 2.45 24.3 5.9
Table 3. Performance of Image-Free Pose Estimation on Our Self-Created Human Dataset with Different Extraction Methods including SPCPose-M1, SPCPose-M2, and SPCPose-M3, Evaluated in Terms of AP, PCK, AUC, EPE, Params, and GFLOPs
View tableView in ArticleTable 3. Performance of Image-Free Pose Estimation on Our Self-Created Human Dataset with Different Extraction Methods including SPCPose-M1, SPCPose-M2, and SPCPose-M3, Evaluated in Terms of AP, PCK, AUC, EPE, Params, and GFLOPs
Method Sample rate Backbone AP@50-95 PCK@0.05 AUC EPE Params (M) GFLOPs SPCPose-M1 4.006 × 10−2 ViT-S 58.0 72.9 56.1 15.97 24.3 5.9 1.002 × 10−2 ViT-S 37.2 55.7 43.0 24.04 24.3 5.9 2.504 × 10−3 ViT-S 27.3 47.7 37.0 27.92 24.3 5.9 6.260 × 10−4 ViT-S 47.3 64.3 50.1 19.05 24.3 5.9 SPCPose-M2 1.002 × 10−2 ViT-S 68.4 78.5 62.7 12.96 24.3 5.9 2.504 × 10−3 ViT-S 63.5 73.7 59.0 14.06 24.3 5.9 6.260 × 10−4 ViT-S 41.1 48.7 45.4 22.18 24.3 5.9 SPCPose-M3 1.002 × 10−2 ViT-S 70.4 80.1 63.2 12.52 24.3 5.9 2.504 × 10−3 ViT-S 69.2 79.5 62.6 12.50 24.3 5.9 6.260 × 10−4 ViT-S 56.3 69.7 54.5 15.87 24.3 5.9
Tools
Get Citation
Copy Citation Text
Xin Wu, Cheng Zhou, Binyu Li, Jipeng Huang, Yanli Meng, Lijun Song, Shensheng Han, "Image-free cross-species pose estimation via an ultra-low sampling rate single-pixel camera," Chin. Opt. Lett. 23, 091101 (2025)
Paper Information
Category: Imaging Systems and Image Processing
Received: Apr. 1, 2025
Accepted: May. 9, 2025
Published Online: Aug. 22, 2025
The Author Email: Cheng Zhou (zhoucheng91210@163.com), Jipeng Huang (huangjp848@nenu.edu.cn), Lijun Song (ccdxslj@126.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20