CCTV4 Feedback #97

Curid commented 2024年08月16日 21:25:10 +02:00

Do you notice any improvement in accuracy both day and night compared to other models?

Is the slower inference speed acceptable, and should the next model sacrifice more speed for better accuracy?

model files

Do you notice any improvement in accuracy both day and night compared to other models? Is the slower inference speed acceptable, and should the next model sacrifice more speed for better accuracy? [model files](https://codeberg.org/Curid/TF-CCTV/src/branch/master/models/cctv4)

harryqt commented 2025年05月31日 10:37:53 +02:00

I was getting too many false positive on the default model provided by Frigate. Google brought me here, and I decided to give cctv4 a shot but still getting false positive.. ☹️

I was getting too many false positive on the default model provided by Frigate. Google brought me here, and I decided to give cctv4 a shot but still getting false positive.. ☹️

Curid commented 2025年05月31日 17:59:37 +02:00

Give yolov9s a try, you'll just have to apply this patch to Frigate.

Give [yolov9s](https://codeberg.org/Curid/TF-CCTV/src/branch/master/yolo/models_official/yolov9s) a try, you'll just have to apply [this](https://codeberg.org/Curid/TF-CCTV/pulls/102) patch to Frigate.

harryqt commented 2025年06月06日 09:37:58 +02:00

@Curid I'm using Coral (usb), is that supported? Frigate documents shows no.

@Curid I'm using Coral (usb), is that supported? Frigate documents shows no.

Curid commented 2025年06月06日 09:55:34 +02:00

Should be

Models ending in "_edgetpu.tflite" are compiled for the Google Coral EdgeTPU. Models without "edgetpu" (simply ".tflite") are for CPU use.

Should be > Models ending in "_edgetpu.tflite" are compiled for the Google Coral EdgeTPU. Models without "edgetpu" (simply ".tflite") are for CPU use.

harryqt commented 2025年06月06日 21:08:58 +02:00

I've been using your recommended model, and I longer get dogs recognized at person even at night, yay..! 🎉 thank you so much.

I've been using your recommended model, and I longer get dogs recognized at person even at night, yay..! 🎉 thank you so much.

harryqt commented 2025年06月22日 06:51:04 +02:00

Reminded me the saying: "never judge too quickly.." 🥲

Reminded me the saying: "never judge too quickly.." 🥲

harryqt commented 2025年09月13日 07:02:50 +02:00

@Curid I’m getting too many false positives — is there a way to fine-tune the YOLOv9s model? Any tutorial or guide you’d recommend?

@Curid I’m getting too many false positives — is there a way to fine-tune the YOLOv9s model? Any tutorial or guide you’d recommend?

Curid commented 2025年09月13日 09:40:42 +02:00

We have a few training script on the temp branch that are ready to run if someone with a good GPU wants to do it.

There's more information about it in the Matrix room

We have a few training script on the [temp branch](https://codeberg.org/Curid/TF-CCTV/src/branch/temp/yolo/models) that are ready to run if someone with a good GPU wants to do it. There's more information about it in the [Matrix room](https://app.element.io/#/room/#tf-cctv:matrix.org)