This repo is a strong Pytorch implementation of SSD ~
- Update the nms to torchvision cuda one, which is much fasater
- Update model init to xavier_uniform_, making the model converge more faster
- Warmup lr is added
- Good performance
method | train data | test data | mAP | aero | bike | bird | boat | bottle | bus | car | cat | chair | cow | table | dog | horse | mbike | person | plant | sheep | sofa | train | tv |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
SSD (Liu et al.) | VOC 2007 trainval | VOC 2007 test | 0.680 | 0.734 | 0.775 | 0.641 | 0.590 | 0.389 | 0.752 | 0.808 | 0.785 | 0.460 | 0.678 | 0.692 | 0.766 | 0.821 | 0.770 | 0.725 | 0.412 | 0.642 | 0.691 | 0.780 | 0.685 |
SSD (mmdet) | VOC 2007 trainval | VOC 2007 test | 0.709 | 0.765 | 0.798 | 0.669 | 0.628 | 0.413 | 0.800 | 0.820 | 0.801 | 0.513 | 0.735 | 0.669 | 0.803 | 0.829 | 0.788 | 0.742 | 0.438 | 0.674 | 0.715 | 0.833 | 0.710 |
SSD (ours) | VOC 2007 trainval | VOC 2007 test | 0.7240 | 0.7761 | 0.8111 | 0.6898 | 0.6253 | 0.4399 | 0.8108 | 0.8368 | 0.8259 | 0.5456 | 0.7754 | 0.7227 | 0.7991 | 0.8242 | 0.7956 | 0.7679 | 0.4565 | 0.6967 | 0.7368 | 0.8300 | 0.7228 |
- The model can be trained smoothly
- The cfg paras is defined in yaml file, which is then merged with args.opts
Wei Liu, et al. "SSD: Single Shot MultiBox Detector." ECCV2016.