【行为检测】G-TAD复现（CVPR 2020）

待我称王封你为后i 2023-01-11 08:48 124阅读 0赞

其实作者给出的环境很详细。这里只是为了记录下环境。便于以后傻瓜复现。

代码参考：

[https://github.com/Frostinassiky/gtad][https_github.com_Frostinassiky_gtad]

论文参考：

[https://arxiv.org/pdf/1911.11462.pdf][https_arxiv.org_pdf_1911.11462.pdf]

# 1.环境 #

ubuntu16.04
    cuda9.0
    python3.6
    torch==1.1.0
    pandas
    h5py
    joblib

安装

cd gtad_lib
    python setup.py install

![20210126105926354.png][]

python align.py

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

# 2.准备数据 #

[THUMOS14][]，下载好后，放到data/thumos\_feature/TSN\_pretrain\_avepool\_allfrms\_hdf5/目录下。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

# 3.训练、推理、后处理 #

# shell script
    python gtad_train.py
    python gtad_inference.py
    python gtad_postprocessing.py
    
    # or shell script
    bash gtad_thumos.sh | tee log.txt

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2][]

经过上述训练、推理、后处理后的结果：

![20210127093458484.png][]

与作者源码中给的结果对比：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3][]

发现结果还是比较好的（一致性高）。

[https_github.com_Frostinassiky_gtad]: https://github.com/Frostinassiky/gtad
[https_arxiv.org_pdf_1911.11462.pdf]: https://arxiv.org/pdf/1911.11462.pdf
[20210126105926354.png]: /images/20221119/f429c93ea2944f6b9c3c140dab7186a5.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221119/8b1746aad1e54e42a921916c0acccd36.png
[THUMOS14]: https://kaust-my.sharepoint.com/personal/xum_kaust_edu_sa/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fxum_kaust_edu_sa%2FDocuments%2FG-TAD&originalPath=aHR0cHM6Ly9rYXVzdC1teS5zaGFyZXBvaW50LmNvbS86ZjovZy9wZXJzb25hbC94dW1fa2F1c3RfZWR1X3NhL0VnVHd3VUdmME8xS3VnX0E2eW0teV84QmxFSjA0X3hQTUU5RUZiQUFLUlBRTnc_cnRpbWU9eHRnb1c2TEIyRWc
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20210126144117896.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20210126211123636.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3,size_16,color_FFFFFF,t_70
[20210127093458484.png]: https://img-blog.csdnimg.cn/20210127093458484.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20210127093708266.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3,size_16,color_FFFFFF,t_70