Approach to collect data

[mobileonekh]

Hi, I am new to YOLO. I have a few question regarding your approach to collect data.

1-Why you capture screen(static image) and count people rather than analyze video instead? I am not sure whether analyze video or static image is more accurate. 2-Is your approach fast enough to analyze live video as it takes time to anaylize each image? 3-You don't set ROI, so same people will appear on multiple snapshot images. How to handle that duplicate?

Sorry, as mentioned about i am new to this Machine learning thing so don't me if my doubt sound ridiculous.

[brian-yu]

Hi @mobileonekh ! Sorry for the late response but here are what I can say about your questions.

1. I capture static images instead of analyzing video since the machine I was running YOLO on was not fast enough to run on a video stream.
2. Yes, YOLO is very fast-- if you have a decent GPU you should be able to run it on every frame of a video stream. According to YOLO's site, "On a Pascal Titan X it processes images at 30 FPS and has a mAP of 57.9% on COCO test-dev."
3. Handling duplicates was outside of the scope that I set for the project. To handle duplicates you may want to investigate other DL architectures.

[guruvishnuvardan]

Hi Brian,

Can you please let me know, If I can use Cmd in the following way

    cmd = "/home/ioz/Desktop/darknet/cfg/coco.data  /home/ioz/Desktop/darknet/cfg/yolov2.cfg /home/ioz/Desktop/yolo.2.0.weights cam.png"#optional flag: -thresh .2"

Can you please correct me, If I am wrong.

Thanks Guru

[brian-yu]

Hey Guru,

Please make sure that the cmd can run in terminal and is a valid executable.

Brian