Use videostream with 1280 resolution

WongKinYiu / yolor

implementation of paper - You Only Learn One Representation: Unified Network for Multiple Tasks (https://arxiv.org/abs/2105.04206)

GNU General Public License v3.0

1.98k stars 524 forks source link

Use videostream with 1280 resolution #202

Open rusvagzur opened 2 years ago

rusvagzur commented 2 years ago

Hi @WongKinYiu,

I observed that inference from videostreams does not work using 1280 resolution. This issue was also addressed here: #62 , where the proposed solution was to use 640 as image resolution. However, your APtest/val results on your tables for d6/w6/p6 configs are based on 1280 resolution.

Can you provide your APtest/val for those models (d6/w6/p6) when using 640?

Are you going to fix this so that 1280 can be used for videostreams?

Thanks in advance

Splendon commented 2 years ago

You can train yolor with img-size of 640 on yourself on coco2017. If train with 8 cards A100，it needs about 4 days to reproduce the result of yolor. Or, you can test the pretrained models on test-dev2017 with img-size of 640. It also can get nearly result. According to my training experience, I think yolor-p6/w6/e6/d6 is not good enough to run with img-size 640, maybe map will lower than yolor-s.