Author Affiliations
1College of Information Science and Technology, Donghua University, Shanghai 201620, China2Engineering Research Center of Digitized Textile & Fashion Technology, Ministry of Education, Donghua University, Shanghai 201620, China;show less
Fig. 1. Improved F-PointNet structure
Fig. 2. Network structure for extracting candidate regions of frustum point cloud
Fig. 3. Registration results of 2D images and 3D point clouds. (a) RGB image; (b) 3D point cloud data; (c) registration effect of Fig. (a) and Fig. (b)
Fig. 4. 3D target frustum candidate region initially obtained
Fig. 5. Schematic of viewing frustum orientation adjustment
Fig. 6. 3D target mask prediction network
Fig. 7. Attention mechanism implementation process
Fig. 8. 3D target bounding box prediction network
Fig. 9. Coordinate transformation of target instance point cloud
Fig. 10. Visual 3D target bounding box prediction results. (a) 2D target detection result; (b) 3D target detection result
Item | CPU | Computing memory | GPU | System | CUDA |
---|
Content | Intel i5-6600 | 8 GB | NVIDIA GTX 1070 | Ubuntu 16.04 | CUDA 9.0 |
|
Table 1. Experimental configuration
xmargin | Car | Pedestrian | Cyclist |
---|
Easy | Moderate | Hard | Easy | Moderate | Hard | Easy | Moderate | Hard |
---|
0 | 82.05 | 68.46 | 62.42 | 65.94 | 58.35 | 50.87 | 74.10 | 55.54 | 52.09 | 0.1 | 82.39 | 69.53 | 62.52 | 61.90 | 55.20 | 49.02 | 73.45 | 55.46 | 52.26 | 0.2 | 82.79 | 70.85 | 63.49 | 67.05 | 59.16 | 51.82 | 76.04 | 57.09 | 53.33 | 0.3 | 83.19 | 70.59 | 63.13 | 65.06 | 57.53 | 50.59 | 73.55 | 55.76 | 52.73 |
|
Table 2. AP values of 3D target detection under each threshold unit: %
Part | AP /% |
---|
Wide-threshold mask(xmargin=0.2) | Attention mechanism | Focal Loss | Easy | Moderate | Hard | - | - | - | 82.05 | 68.46 | 62.42 | √ | - | - | 82.79 | 70.85 | 63.49 | - | √ | - | 81.89 | 69.23 | 62.54 | - | - | √ | 82.73 | 69.89 | 63.27 | √ | √ | √ | 83.04 | 71.25 | 63.82 |
|
Table 3. Influence of each processing part on AP values
Method | AP /% |
---|
Easy | Moderate | Hard |
---|
MV3D[4] | 71.29 | 62.28 | 56.56 | F-PointNet[5] | 82.05 | 68.46 | 62.42 | UberATG-ContFuse[14] | 82.54 | 66.22 | 64.04 | MLOD[15] | 72.24 | 64.20 | 57.20 | Proposed | 83.04 | 71.25 | 63.82 |
|
Table 4. Comparison of AP values of different models