search

shihenw / convolutional-pose-machines-release

Code repository for Convolutional Pose Machines, CVPR'16
Other
878 stars 341 forks source link

Network Speed #14

Open tetmin opened 8 years ago

tetmin commented 8 years ago

Hi Shih-En,

Thanks so much for sharing your work. I've implemented your code & it's working great, but I don't get anywhere near realtime performance using a very good GPU. How did you achieve your videos with the realtime performance?

Also, are you using any bayesian filtering for tracking the estimates?

Best regards, Tom

shihenw commented 8 years ago

There might be these difference: 1) Network itself: we just add a new model that runs faster and scores higher on MPI. Please pull repo and run testing/get_model.sh again. 2) We only use single scale testing for the real-time system. You can reduce the number of scales to be search in config.m. For above please see this demo ipython notebook for more details and explanations. 3) We used 4 GPUs (all Titan X) to scale up the throughput. 4) We used CUDA to accelerate other image processing, and result rendering etc, in the real-time system.

Finally, no we didn't use any temporal information in the video. All results shown are per-frame detection.

thuyang commented 8 years ago

Hi Shih-En

Thank you for sharing your great work. 1) I am wondering which model is the fastest for pose estimation? You mentioned in this thread that you do have a faster and more accurate model. an you specify the name? 2) It seems the llink of python notebook is corrupt. Where can I find details regarding the network speed?

Thank you

mindcont commented 7 years ago

Hello everyone, In gpu mode, one nvidia gtx 1080 which has 8 g graphics memory can use to train the network and real-time forecasting?

If anyone knows, please tell me, much thanks!