mmdetection versions of "Learning Rich Features at High-Speed for Single-Shot Object Detection", ICCV, 2019
This is a translation version of the offical implemention vaesl/LRF-Net. You can easily customize your own dataset with the help of mmdetection toolbox.
Please follow the instructions of mmdetection/docs/INSTALL.md. Put my code into the corresponding folders of mmdetection. The tutorial of mmdetection is available at GETTING_STARTED.
| Dataset | This Repository | Offical Version |
|---|---|---|
| coco | √ | √ |
| Pascal VOC | √ | × |
All the following scores are trained with VOC 07+12 trainval by 24 epoch and test in VOC07 test.
| Models | Performance |
|---|---|
| SSD300 | 0.767 |
| LRF300 | 0.797 |
| SSD512 | 0.793 |
| LRF512 | 0.818 |
All of the following scores are from the official trained models which transformed into mmdetection version. i.e. I only change the keys of OrderedDict.
| Models | Train DataSet | Test DataSet | This Repository |
|---|---|---|---|
| LRF300 | coco2014 trainval | coco2017 val | 0.267 |
| LRF512 | coco2014 trainval | coco2017 val | 0.31 |