This is an implementation of multi-scale CenterNet based on FPN. The results are verified on VOC2007.
Here I do not use DCNv2 and deconvolution. The feature fusion method follows a standard FPN. The model is testd on VOC2007. Single-scale model only utilizes feature C2 to predict. Multi-scale model utilizes C2, C3, C4 to predict. In addition, I adopt the training trick in SNIP and Trident Networks. Objects of different scales are assigned to C2, C3 and C4 respectively.
| Model | resolution | AP |
|---|---|---|
| resnet18-dcn | 384 | 72.6 |
| resnet101-dcn | 384 | 77.1 |
| resnet18-fpn-c2 | 384 | 72.46 |
| resnet18-fpn-multi | 384 | 74.97 |
| resnet101-fpn-multi | 384 | 79.16 |