Table 2.
Characteristics comparison between different approaches for car classification. The first group is described in Section 4: All-vs.-all (AVA), Two-step cascade (2SC), and Hierarchical multilabel (HML). The second group is described in Section 2: Channel Max Pooling (CMP), Attention Branch Network (ABN), and the method by Fomin et al.
| Approach | Exploits Hierarchy |
Car Level Prediction |
Single Input |
Number of Classifiers |
Backbone Architecture |
|---|---|---|---|---|---|
| AVA (ours) | ✘ | MMY | ✔ | 1 | ResNet-50 |
| 2SC (ours) | ✔ | Type → MMY | ✔ | 1 + 12 | ResNet-50 |
| HML (ours) | ✔ | Type → MMY | ✔ | 1 | ResNet-50 |
| CMP [18] | ✘ | Model | ✔ | 1 | DenseNet-161, VGG16, ResNet-152 |
| ABN [22] | ✘ | Model | ✔ | 1 | VGG16, ResNet-101 |
| Fomin et al. [23] | ✔ | Orientation → Make | ✘ | 1 + 5 | Inception-ResNet-v3 |