@ideaRunner

[YaaYang]

@ideaRunner After you use TSN to extract video features, the features length is different. You must reshape these features by fitting and interpolation functions.

Originally posted by @lijiannuist in https://github.com/TencentYoutuResearch/ActionDetection-DBG/issues/1#issuecomment-554289967

how extract feature from tsn

[YaaYang]

@ @lijiannuist

[swordlidev]

@YaaYang You must follow https://github.com/yjxiong/temporal-segment-networks for video classification. Then with these classification models, you can extract features by using the FC layer.

[YaaYang]

@YaaYang You must follow https://github.com/yjxiong/temporal-segment-networks for video classification. Then with these classification models, you can extract features by using the FC layer.

thanks 所以说ActionDetect-DBG ReadME中说的rescale 就是说 global pool 之后的特征 连接一个 fc 然后 输出特征长度是100 是吗？？？

[YaaYang]

@lijiannuist 所以就是 使用 fc 之后的特征向量 并不是 global pool 之后的对吧

[swordlidev]

@YaaYang rescale 是额外做的。一个视频抽完特征大小NX400，再rescale到100X400。

[YaaYang]

@lijiannuist

[YaaYang]

@lijiannuist thanks 这里的抽取特征就是指 global pool 之后的特征吗 然后 做 rescale 对吧

[YaaYang]

@lijiannuist Sorry 这里说的 global pool 是 https://github.com/yjxiong/temporal-segment-networks 中 inception FC的前一层 ， 视频抽取特征是这个global pool的输出吗 thanks

[swordlidev]

This layer ,200X2=400, 200 flow feature, 200rgb feature , 200 is also activitynet class number.

[YaaYang]

@lijiannuist Thanks ，，，，，，，，，，，，，，

[ideaRunner]

Hi @YaaYang , I am doing this part. As @lijiannuist is so warm-hearted and release the code of TSN feature extraction, I will not be confused by the parameter in anet2016

I think the feature extraction code is from TSN official project and the model is released from anet2016 TSN project

BTW, I have another question. As we don't need to re-train TSN model and the TSN feature extraction don't need the over_sample method, which caffe version should we use ? the caffe modified and optimized by yjxiong or the official version caffe?

[gtgtgt1117]

Hi @YaaYang , I am doing this part. As @lijiannuist is so warm-hearted and release the code of TSN feature extraction, I will not be confused by the parameter in anet2016

I think the feature extraction code is from TSN official project and the model is released from anet2016 TSN project

BTW, I have another question. As we don't need to re-train TSN model and the TSN feature extraction don't need the over_sample method, which caffe version should we use ? the caffe modified and optimized by yjxiong or the official version caffe?

可能做了一些优化,但是,可以直接编译作者提供的lib/caffe_action里面的caffe就可以,我是这么做的

[ideaRunner]

@gtgtgt1117 Yes, You are right. But there will be a problem when doing inference acceleration. There is a custom op in yjxiong's caffe, and it is not supported by other boost tools.