FFBNET : LIGHTWEIGHT BACKBONE FOR OBJECT DETECTION BASED FEATURE FUSION BLOCK
| System | mAP | FPS (1080Ti) |
|---|---|---|
| Mob-SSD | 68 | 190 |
| Tiny-Yolo v3 | 61.3 | 220 |
| Pelee | 70.9 | - |
| SSD | 77.2 | 160 |
| STDN | 78.1 | 41 |
| FSSD | 78.8 | 150 |
| RefineDet | 80.0 | - |
| FFBNet | 73.54 | 185 |
| VGG-FFB | 80.2 | 142 |
-
Install PyTorch 0.3.1 by selecting your environment on the website and running the appropriate command.
-
Clone this repository. This repository is mainly based onlzx1413/PytorchSSD, and a huge thank to him.
-
Compile the nms and coco tools:
./make.sh# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory># specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>- First download the fc-reduced VGG-16 PyTorch base network weights at: BaiduYun Driver, password is
mu59. - MobileNet is reported in the paper, weight file is available at: BaiduYun Driver, password is
f7oe.
# Put vgg16_reducedfc.pth, and mobilenet_1.pth in a new folder weights and
python train_test_mob.py or python train_test_vgg.pyPersonal advice: when use Mobilenet v1 to train voc datasets, use a higher learning rate at the beginning, the convergence performance may be better.
If you are interested in this paper or interested in lightweight detectors, please QQ me (374873360)