[Detector Question]: which tensorrt model to use?

[vajonam]

Describe the problem you are having

Just curious, with the previous release we had access to yolov4, and now have yolov7. I noticed the the yolov7-320 model had some more false positives. More than the previous yolov4-416 model I was using. Moving to yolov7-640 helped with the false positives, but the inference time on my GPU is up to 70ms from <20ms. My question is, if should I keep with the yolov4-416 model or move over to the yolov7 mode, what are others expereinces here?

[vajonam]

@NateMeyer 320 seems a bit too small for me and my camera resolutions, too many false positives and 640 might be overkill interns of GPU horse power I have. 416 in yolov4 was a good compromise for me, any chance we can get that added back as an option for yolov7?

I have added this locally for my container environment, overriding

- /opt/docker/configs/frigate/override/download_yolo.sh:/usr/local/src/tensorrt_demos/yolo/download_yolo.sh

added

if [[ -z "$TARGET_MODEL" ]] || [[ $TARGET_MODEL =~ .*"yolov7-416"*. ]]; then
    echo "Creating yolov7-416.cfg and yolov7-416.weights"
    cat yolov7.cfg | sed -e '6s/batch=64/batch=1/' | sed -e '8s/width=640/width=416/' | sed -e '9s/height=640/height=416/' > yolov7-416.cfg
    ln -sf yolov7.weights yolov7-416.weights
fi

I will report back on this and how it works for possible inclusion into your repo?

[NateMeyer]

Let me know how it goes. If that works well, we can add that to tensorrt_demos repo.

[vajonam]

So far its been working well for me. However its detection some "squirrels" as person, but that is easily filtered out using min size for a person, but other than that quite reliable, its about 1/2 the inference time, but doesn't seem like half the precision. I dont have a quantitative test setup using existing videos. but that is where I am at with this, would vote to add it.

[vajonam]

the false positives I was getting with shadows and IR flight flashes as persons are now gone. So I would say its helped in that regard without increase inference time too much.

[twister36]

I agree yolov7-320 seems to have allot of false positives. I too am going to try the other models...

I see in the online documentation a list of the models available, and yolov4-416 is available (@vajonam). However do we have a better list that defines each models accuracy vs speed?

[vajonam]

I have not quantified this, but created a yolov7-416 model that I have been using with minimal false positives. Instructions above.

[twister36]

I'll have to give it a try. I did find a few links that describe what the different models do and what they are used for. Right now I'm testing yolo7-640z.

[Sceptersax]

@twister36 how's yolov7-640? Significantly more accurate or just slightly?