mzolfaghari / ECO-efficient-video-understanding

Code and models of paper " ECO: Efficient Convolutional Network for Online Video Understanding", ECCV 2018
MIT License
437 stars 96 forks source link

I have some questions about the speed of I3D to ask you. #51

Closed Jiangchitang closed 5 years ago

Jiangchitang commented 5 years ago

Dear author, I am fortunate to read your thesis. I have some questions about the speed of recognition to ask you. The speed of I3D in your paper is only 0.9 clip/s. but the speed I tested reached 40 clip/s ( Titan V). Is the method we tested different? I hope to know more details.

mzolfaghari commented 5 years ago

@Jiangchitang The reported speed is the speed of video inference "Number of complete videos evaluated by the model". To calculate this you need to consider augmentations of video and splitting it to clips. By clipping and augmentations, the speed decreases to the number we've reported. Of course, if you consider only a single clip and just center cropping the speed will be much higher.

Jiangchitang commented 5 years ago

Thinks!