https://github.com/albert-jin/CvT-SSD

https://github.com/albert-jin/CvT-ASSD

1. You may probably need to install an anaconda environment which contains all packages followed.

   • pytorch 1.9.0 py3.7_cuda10.2_cudnn7_0 pytorch
   • cudatoolkit 10.2.89 h74a9793_1
   • opencv-python 4.5.2.54 pypi_0 pypi
   • visdom 0.1.8.9 pypi_0 pypi
   • yacs 0.1.8 pypi_0 pypi
   • jupyter 1.0.0 pypi_0 pypi

2. For training, an NVIDIA GPU is strongly recommended for speed. we use two NVIDIA GTX-1080TI, but we recommend GPUs like Tesla-V100 /RTX-3090 for more memory

3. Before you run the codes for self-study or reappearance the performance in this paper "CvT-ASSD", please add the CvT_SSD/model/ directory into sources Root caused by the reference of many codes inside of model directory

4. you should download the pytorch parameters file postfix by ".pth" and move into models/CvT/weights like 项目结构.PNG

5. 图像物体检测benchmark(参照论文native-SSD)一般是将VOC2007—TEST的数据作为模型的测试集,训练集可有以下搭配:

   • 1. 07:VOC2007 trainval 训练集验证集
   • 2. 02+12 VOC2007 trainval + VOC2007 trainval 训练集验证集
   • 3. 07+12+COCO 在 COCO trainval35k上预训练,然后在07+12上微调

6. 评价指标maP使用mxnet提供的VOC07MApMetric,将recall分成10等分,继而对所有precision取平均,在对类别去平均,具体参见 https://blog.csdn.net/u014203453/article/details/77598997