[FEATURE] How can i modified the demo.py to inferece the camera?

[JiangWeiHn]

I want to analyze the video data of the camera in real time.Then How do I revise the demo.py to fit this mission?

[ujjawalcse]

Hi @JiangWeiHn, Did you get any way for realtime camera inference? Please share with me in case you got something.

[JiangWeiHn]

Hi @JiangWeiHn, Did you get any way for realtime camera inference? Please share with me in case you got something.

I have try import multiprocess in the demo.py,but it is still run slowly,There's a delay of about three to four seconds。I check the process,the rendering program takes a lot of time.Up to now , I have no idea to improve this part of program

[ikvision]

I think SPIN is a better fit for this application as it operates frame level. SPIN delay will be the processing time of the backbone of a single image, not of sequence of frames

[ujjawalcse]

What if we don't render it on SMPL mesh and just extract the corresponding 3D joints to animate a predefined skeleton or a rig (simple edge-based skeleton) in realtime. I think it would take less time and also useful for animated file generation that can be used in any 3D tools like blender.

[ikvision]

Most of the compute time per frame is spent in the backbone (resnet50 in HMR), therefore removing the mesh generation (in SMPL model) is not likely to get a significant speedup. The fact that VIBE uses multiple frames (GRU over tone) is probably the major latency factor compared to SPIN.

[JiangWeiHn]

thanks a lot ,I will try SPIN.if there who got something, please sharing us in here

[Pranjal2041]

I have opened a pr for inferencing from camera using VIBE, and it works well on my system(gtx1060 6gb). Inference speed is roughly 15 frames per second for sequence length of 4 and yolo_img_size of 256. Change in sequence length doesn't change the speed a lot since only the encoder and regressor part needs to be done on the whole sequence. Though this is without the rendered results being displayed in real-time(It's way too slow), but almost all other features are intact. Do let me know if any other changes are needed.

Hope it Helps. #40

[EvgeniaKomleva]

@Pranjal2041 Please explain how you launched the project online? I could not find flags or any other instructions

[Pranjal2041]

Hi @EvgeniaKomleva, I had created a new file 'live_inference.py.' You can run that for online inference.