Fig. 1. The structure of our proposed 3D concealed object detector for AMMW holographic images
Fig. 2. Projection of the AMMW holographic image (a) AMMW holographic image, (b) the resulting 2D front view of performing projection along the Z axis of the holographic image in Fig. 2(a), (c) the shape and size changes caused by projecting a 3D object into 2D views
Fig. 3. The structure of our proposed input module
Fig. 4. The distribution of the number of points in the bounding box in 3D point clouds and 2D front images
Fig. 5. The structure of our proposed 3D feature extractor
Fig. 6. The structure of our proposed output module
Fig. 7. Comparison of localization and detection performance for different networks (a) Recall as a function of IOU threshold, (b) PR curve under IOU = 0.5
Fig. 8. Qualitative detection results of different networks, where the red bounding boxes denote the ground-truth, and the yellow bounding boxes denote the predicted results (a)-(d) Our proposed method, (e) RPN, (f) Faster RCNN, (g) RetinaNet, (h) TridentNet
Fig. 9. F1-score under different thresholds of the confidence
网络组件 | 卷积类型 | 通道 | 卷积核 | 步长 | 填充 | 空洞率 | 感受野 | 特征图 |
---|
降采样模块 | SubMConv | 16 | 3 | 1 | 1 | 1 | 3 | (192,400,41) | SubMConv | 16 | 3 | 1 | 1 | 1 | 5 | (192,400,41) | SpConv | 32 | 3 | 2 | 1 | 1 | 7 | (96,200,21) | SubMConv | 32 | 3 | 1 | 1 | 1 | 11 | (96,200,21) | SpConv | 64 | 3 | 2 | 1 | 1 | 15 | (48,100,11) | SubMConv | 64 | 3 | 1 | 1 | 1 | 23 | (48,100,11) | SpConv | 128 | 3 | (1,1,2) | (1,1,0) | 1 | 31 | (48,100,5) | 分支1 | SubMConv | 128 | 3 | 1 | 1 | 1 | 39 | (48,100,5) | SpConv | 128 | (1,1,3) | (1,1,2) | 0 | 1 | 39 | (48,100,2) | 分支2 | SubMConv | 128 | 3 | 1 | (2,2,1) | (2,2,1) | 47 | (48,100,5) | SpConv | 128 | (1,1,3) | (1,1,2) | 0 | 1 | 47 | (48,100,2) | 分支3 | SubMConv | 128 | 3 | 1 | (3,3,1) | (3,3,1) | 55 | (48,100,5) | SpConv | 128 | (1,1,3) | (1,1,2) | 0 | 1 | 55 | (48,100,2) |
|
Table 1. Details of convolutional layers of the 3D feature extractor
网络结构 | 上下文信息提取模块 | X、Y方向降采样步长 | AP50 |
---|
SECOND[11] | × | 8 | 72.28 | SECOND + HRF | × | 4 | 73.87 | SECOND + CIE | √ | 8 | 73.85 | SECOND + HRF + CIE | √ | 4 | 74.44 |
|
Table 2. Comparison of the AP with different network structures
网络 | 输入 | AP10 | AP20 | AP30 | AP40 | AP50 | 平均AP | FA50 | Re50 | 速度(FPS) |
---|
RPN[7] | 2D | 82.35 | 79.26 | 74.15 | 64.26 | 48.50 | 69.70 | 36.68 | 57.99 | 23.0 | Faster RCNN[7] | 2D | 88.88 | 87.53 | 84.76 | 78.71 | 67.33 | 81.44 | 22.74 | 67.51 | 4.7 | RetinaNet[14] | 2D | 89.65 | 88.76 | 86.20 | 80.11 | 65.44 | 82.03 | 24.59 | 67.18 | 9.0 | TridentNet[17] | 2D | 91.38 | 89.90 | 87.07 | 80.10 | 64.87 | 82.66 | 23.52 | 67.04 | 4.1 | SECOND[11] | 3D | 92.56 | 91.75 | 89.87 | 84.97 | 72.28 | 86.29 | 21.24 | 74.65 | 22.9 | 本工作 | 3D | 92.97 | 92.04 | 90.02 | 85.34 | 74.44 | 86.96 | 20.96 | 76.26 | 17.3 |
|
Table 3. Comparison of detection performance on different Networks