search

kundtx / lfd2022-comments

0 stars 0 forks source link

Learning from Data (Fall 2022) #15

Open kundtx opened 1 year ago

kundtx commented 1 year ago

http://8.129.175.102/lfd2022fall-poster-session/11.html

yang2011 commented 1 year ago

（Instructor）Well written poster. What does "Anchor-only" mean in the experiment? How is an anchor frame defined?

ccp123456 commented 1 year ago

G40 ChuPeng: Excellent project! I want to know the time cost of feature extraction and the detection delay?

Wanghaha2333 commented 1 year ago

@yang2011 （Instructor）Well written poster. What does "Anchor-only" mean in the experiment? How is an anchor frame defined?

G11 Fulin Wang Dear professor Li, Thanks for your question! Anchor frames means that only the first frame of every four frame's object detection result is reused on the next three frames. In our method Safecare, besides the anchor frames, the non-anchor frames with SSIM<0.87 will also be preserved to be inferred by YoloV5 model, which guarantees that the video frames with significant changes will not be filtered. (Due to the lack of the poster's space, I intend to present this orally.)

Wanghaha2333 commented 1 year ago

@ccp123456 G40 ChuPeng: Excellent project! I want to know the time cost of feature extraction and the detection delay?

G11 Fulin Wang Dear Chupeng, Thanks for your question! The speed of feature extraction on our server is about 74fps (which means the time cost of extracting the difference feature from two adjacent frames is 1/74 seconds) on 720p video frames. And the detection speed of YoloV5 is about 10fps, which is much smaller than the video's frame rate. Therefore, filter some unimportant video frames could help to boost the processing of the whole video.

yuyan12138 commented 1 year ago

G29 Yuyan Wang: Good work! I wonder is there any other way to speed up the video analysis process besides the approach in this article?

lpwpower commented 1 year ago

@yuyan12138 G29 Yuyan Wang: Good work! I wonder is there any other way to speed up the video analysis process besides the approach in this article? G11 Peiwen Li Thanks for your question, Yuyan! For this problem, we are also considering the use of model lightweighting methods to solve, for example reducing the size of the model by incorporating the removal of the inadequate components and thus produce a pruned lightweight model.