search

Shared-Reality-Lab / IMAGE-server

IMAGE project server components
Other
2 stars 7 forks source link

Upgrading to YOLOv6 and using larger YOLO models #209

Closed JRegimbal closed 2 years ago

JRegimbal commented 2 years ago

Based on https://github.com/Shared-Reality-Lab/auditory-haptic-graphics-server/blob/e77d7c39cd589836bec69eb06c28920ce1772fba/preprocessors/yolo/detect.py#L122 it seems like we're using YOLOv5s. There are a lot of other YOLO models with more parameters that may result in better accuracy for reasonable increases in resource consumption and time. If tests have been done, can this data be put in the Wiki or another public space? If tests have not been done, can we please do them and make a determination of which model is best for our needs?

rohanakut commented 2 years ago

@JRegimbal I had seen the remaining models. The entire spectrum of YOLOv5 models can be found here. I haven't tested all the models, but according to my experience, to see a significant change in the accuracy we might have to go with YOLOv5x or higher. It might result in a significant delay in the results as anything beyond v5x takes atleast 4sec to run on v100(32gb) GPU(not really sure how long it would take on bach).

JRegimbal commented 2 years ago

Where are you getting 4s from? The speeds in the table are reported in milliseconds.

jeffbl commented 2 years ago

@rohanakut @gp1702 is there a holdup on switching to the larger YOLO model? Have you encountered issues when attempting to doing so?

rohanakut commented 2 years ago

@jeffbl i havent tried testing the model until now. I have 2 report submissions due today. So the earliest I can test these models is tomorrow.

jeffbl commented 2 years ago

@rohanakut did you communicate this to @gp1702, so he could potentially take over?

rohanakut commented 2 years ago

yes i have communicated the same with @gp1702 . @gp1702 you might wanna discuss this with jeff

gp1702 commented 2 years ago

Unfortunately, I cannot take over today. I have rebuttal submission due at 9.

On 07-Dec-2021, at 2:35 PM, rohanakut @.***> wrote:

 yes i have communicated the same with @gp1702 . @gp1702 you might wanna discuss this with jeff

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

rohanakut commented 2 years ago

The model weights of the larger YOLO model are incompatible with the current version. This is a known issue, for more info see #5175 in the YOLO repository.

We can fix this by upgrading to YOLOv6, and this would resolve #208. But to do that I need to make a few changes in detect.py, and this might take some time.

jeffbl commented 2 years ago

I believe #229 should close this? Please make sure to link issues to their PRs.

rohanakut commented 2 years ago

@jeffbl closing this since issue is resolved