Infrared and Laser Engineering, Volume. 53, Issue 8, 20240206(2024)
Semi-supervised 3D object detection based on frustum transformation and RGB voxel grid
Figures & Tables(18)
Fig. 1. A 3D object detection framework based on frustum transformation and RGB voxel maps
Fig. 2. RGB voxel feature extraction module based on frustum transformation
Fig. 6. The RVFM addresses the directional issue, as shown in the bottom left and right figures
Fig. 7. The RVFM addresses the proximity issue, with the results displayed in the bottom left and right figures
Fig. 8. Comparison of visualization results between the CAM module and other fusion modules
Fig. 9. The successful scenarios demonstrate high detection accuracy for both cars and cyclists
Fig. 10. The successful scenarios demonstrate high detection accuracy for occluded objects
Fig. 11. The CLM module resolved the false detection issues, as shown in the bottom left and right images
Fig. 12. The complete model addressed issues of repetitive detections and missed detections
Fig. 13. The complete model resolved false detection issues and detected unlabelled objects
Table 1. Performance comparisons with established methods were conducted on the KITTI test set
View tableView in ArticleTable 1. Performance comparisons with established methods were conducted on the KITTI test set
Method Speed Car 3D AP(R40) Pedestrian 3D AP(R40) Cyclist 3D AP(R40) Easy Moderate Hard Easy Moderate Hard Easy Moderate Hard PointPillars[ 14 ]61.2 79.05 74.99 68.30 52.08 43.53 41.49 75.78 59.07 52.92 VoxelNet[ 2 ]4.5 81.97 65.46 62.85 57.86 53.42 48.87 67.17 47.65 45.11 SECOND[ 15 ]20 83.13 73.66 66.20 51.07 42.56 37.29 70.51 53.85 46.90 MV3D[ 3 ]2.8 71.09 62.35 55.12 - - - - - - UberATG-ContFuse[ 16 ]16.8 82.54 66.22 64.04 - - - - - - F-PointNet[ 4 ]6 81.20 70.39 62.19 51.21 44.89 40.23 71.96 56.77 50.39 AVOD-FPN[ 17 ]12.5 81.94 71.88 66.38 50.80 42.81 40.88 64.00 52.18 46.61 MVXNet[ 18 ]25.2 83.20 72.70 65.20 - - - - - - PointPainting[ 19 ]28.5 82.11 71.70 67.08 50.32 40.97 37.77 77.63 63.78 55.89 Ours(LRFN) 21 83.26 74.51 69.76 48.08 40.21 37.83 71.55 58.15 51.94 Baseline-VoxelNet 4.4 80.74 65.58 63.73 56.44 52.16 47.79 66.39 45.81 44.37 Ours(LRFN-S) 21.4 88.12 75.88 71.21 59.67 56.30 52.61 72.57 56.50 52.74
Table 2. The influence of each module in the LRFN model on the performance of the KITTI validation set
View tableView in ArticleTable 2. The influence of each module in the LRFN model on the performance of the KITTI validation set
Method RVFM CAM AP 3D Easy Mod. Hard Ours - - 69.93 59.44 56.81 √ - 70.49 60.17 58.07 √ √ 74.41 62.65 59.66 Improvements +4.48 +3.21 +2.85
Table 3. The impact of different fusion modules on the performance of the KITTI validation set
View tableView in ArticleTable 3. The impact of different fusion modules on the performance of the KITTI validation set
Method AP 3D mAP Easy Moderate Hard VC[ 18 ]63.32 50.39 47.23 53.65% LI-F[ 20 ]70.93 59.50 56.84 62.42% Ours(CAM) 74.41 62.65 59.66 65.57%
Table 4. Ablation study on 1% labeled data from the KITTI dataset
View tableView in ArticleTable 4. Ablation study on 1% labeled data from the KITTI dataset
Method Car 3D AP(R40) Pedestrian 3D AP(R40) Cyclist 3D AP(R40) Easy Moderate Hard Easy Moderate Hard Easy Moderate Hard PV-RCNN[ 6 ]87.7 73.5 67.7 32.4 28.7 26.2 48.1 28.4 27.1 naive psd.-lb.[ 5 ]88.4 75.2 69.5 32.7 29.2 26.7 51.4 30.7 28.7 cls. filt. only[ 5 ]87.9 75.5 70.5 36.6 31.0 28.3 57.3 35.3 33.0 3DIoUMatch[ 5 ]89.0 76.0 70.8 37.0 31.7 29.1 60.4 36.4 34.3 Ours 90.7 77.5 72.2 39.1 31.8 31.9 60.9 38.7 35.5
Table 5. Ablation study of the effectiveness of different modules
View tableView in ArticleTable 5. Ablation study of the effectiveness of different modules
Method CLM SFF GBV Car-3D Detection mAP Easy Mod Hard Baseline - - - 81.80 66.92 63.97 70.90% Ours - √ - 83.24 68.21 66.50 72.65% √ √ - 86.54 75.78 73.57 78.63% √ √ √ 87.36 76.71 74.46 79.51%
Tools
Get Citation
Copy Citation Text
Yan WANG, Tiantian YUAN, Bin HU, Yao LI. Semi-supervised 3D object detection based on frustum transformation and RGB voxel grid[J]. Infrared and Laser Engineering, 2024, 53(8): 20240206
Paper Information
Category:
Received: May. 16, 2024
Accepted: --
Published Online: Oct. 29, 2024
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20