直接运行main_cla.py即可,train.py里的参数修改后会覆盖main_cla.py
主要修改的地方为:
(1)models/deepSBD.py
将网络结构修改为deepSBD 中介绍的模式
(2)opts.py
根据实际情况修改部分参数,解决了源代码的报错情况
(3)split_data.py
由于数据集过大,解压移动时造成了部分数据的损失,因此从原数据集中随机选择了100个视频,所有以choose开头的文件都是后面随机选择的。数据不对应会导致程序无法正常运行。 由于只是用于实验,该代码还不完善,如果使用完整数据集则不需要运行此代码。
(4)eval_res.py
调整了输出格式使其规划化。
This repository contains our implementation of deepSBD for ClipShots. The code is modified from here.
We implement deepSBD in this repository. There are 2 backbones that can be selected, including the original Alexnet-like and ResNet-18 introduced in our paper.
- The trained model for Alexnet-like backbone. BaiduYun, Google Drive
- The trained model for ResNet-18 backbone. BaiduYun, Google Drive
- The pretrained model for ResNet-18 backbone.
Please refer to train.sh
Add '--no_train' options to train.sh
.