Figure 1:
Schematic design of the proposed modified CenterNet model.
Figure 2:
ResNet50 architecture.
Figure 3:
ResNext50 architecture.
Figure 4:
Inception-ResNetV2 architecture.
Figure 5:
DenseNet201 architecture.
Figure 6:
Original images with ground truth labels.
Figure 7:
Comparative analysis of training loss.
Table 1:
Comparison of training loss
| Model | Regression loss | Mask loss | Total loss |
|---|---|---|---|
| Center-Inception-ResNetV2 | 0.85244 | 1.6311645 | 2.484085 |
| Center-ResNet50 | 0.766475 | 0.666525 | 1.432925 |
| Center-DenseNet201 | 0.450010 | 2.514673 | 2.964682 |
| Center-ResNext50 | 0.466475 | 0.266525 | 0.733000 |
Table 2:
Performance evaluation of proposed models based on A3DP-Abs
| Model | Mean | C-l (loose criteria) | C-S (strict criteria) |
|---|---|---|---|
| Center-DenseNet201 | 39.92 | 52.284 | 44.68 |
| Center-ResNet50 | 38.854 | 50.120 | 43.44 |
| Center-Inception-ResNetV2 | 38.399 | 48.614 | 43.20 |
| Center-ResNext50 | 37.06 | 47.027 | 41.52 |
Table 3:
Performance evaluation of SOTA models based on A3DP-Abs
| Model | Mean | l (loose criteria) | S (strict criteria) |
|---|---|---|---|
| 3D-RCNN (CVPR,18) | 16.44 | 29.70 | 19.80 |
| Direct-Based (CVPR,19) | 15.15 | 28.71 | 17.82 |
Table 4:
Performance evaluation of proposed models based on A3DP-Rel
| Model | Mean | C-l (loose criteria) | C-S (strict criteria) |
|---|---|---|---|
| Center-DenseNet201 | 11.82 | 23.89 | 10.85 |
| Center-ResNet50 | 10.51 | 23.00 | 9.50 |
| Center-Inception-ResNetV2 | 9.81 | 22.60 | 9.42 |
| Center-ResNext50 | 9.61 | 22.17 | 9.04 |
Table 5:
Performance evaluation of SOTA models based on A3DP-Rel
| Model | Mean | C-l (loose criteria) | C-S (strict criteria) |
|---|---|---|---|
| 3D-RCNN (CVPR,18) | 10.79 | 17.82 | 11.88 |
| Direct-Based (CVPR,19) | 11.49 | 17.82 | 11.88 |
Figure 8:
Comparative analysis based on A3DP-Abs. A3DP-Abs, absolute translation thresholds.
Figure 9:
Comparative analysis based on A3DP-Rel. A3DP-Rel, relative translation thresholds.