Issue regarding generating blob file

[drakshayanidesai]

I tried with hub.ultralytics with VisDrone dataset with yolovV5m6. It is done but it is throwing error while converting to blob files using tools.luxonis.com(The sheet attached shows the errors). Then I tried with yolov5n model on the same dataset but it is showing error while executing in google colab notebook. Please see the errors in the attached sheet. Suggestions?

Regards Drakshayani Luxonis.docx

[tersekmatija]

@HonzaCuhel can you look into this? Might be related to TOOLS-LUXONIS-2N Sentry issue?

@drakshayanidesai Can you share .pt weights with us please so we can debug the issue further?

[drakshayanidesai]

I sent you through mail

[tersekmatija]

@drakshayanidesai Sorry, which email?

[cafemoloko]

@drakshayanidesai, could you please send an email to support@luxonis.com? Thank you!

[HonzaCuhel]

Hello @drakshayanidesai,

I investigated why the conversion of the yolovV5m6 fails. I found out that the image size 640x352 is not a multiple of maximum stride 64 (because 352 is not a multiple of 64), which until now was not shown in the error message in the tools. I apologize for the inconvenience. I have prepared an update for the tools that fixes that. This update will be deployed in the following days. So my suggestions to you are either, please try to export the model with a different image size (e.g., 640x320 or 640x384) or use a different model, such as yoloV5n or yoloV5s, which both will run faster on OAK devices than yolovV5m6 and both can be converted with image size 640x352.

Best Jan

[drakshayanidesai]

Hi Jan

I tried yoloV5s with VisDrone dataset. But Google colab notebook is running out of time each time I am trying. So could you please send the trained model for the same (blob).It is very much in need.

Regards Drakshayani On Sat, 18 Mar 2023 04:29:04 +0530 Honza Čuhel wrote

Hello @drakshayanidesai, I investigated why the conversion of the yolovV5m6 fails. I found out that the image size 640x352 is not a multiple of maximum stride 64 (because 352 is not a multiple of 64), which until now was not shown in the error message in the tools. I apologize for the inconvenience. I have prepared an update for the tools that fixes that. This update will be deployed in the following days. So my suggestions to you are either, please try to export the model with a different image size (e.g., 640x320 or 640x384) or use a different model, such as yoloV5n or yoloV5s, which both will run faster on OAK devices than yolovV5m6 and both can be converted with image size 640x352. Best

Jan

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[HonzaCuhel]

Hi @drakshayanidesai,

I am not sure if I understand you correctly, for the investigation described above I have used the standard yolov6m6 pretrained on the COCO dataset (I haven't trained it or any other model on the VisDrone dataset). So, which model/blob did you exactly mean?

Best Jan

[drakshayanidesai]

Hi Jan

Actually my task is counting number of cyclist, pedestrian in real time using OAK-D. So I want to use VisDrone dataset for training model(yoloV5s). Hence trying in colab not able to achieve hence asked you.

Regards Drakshayani

On Mon, 20 Mar 2023 14:08:03 +0530 Honza Čuhel wrote

Hi @drakshayanidesai, I am not sure if I understand you correctly, for the investigation described above I have used the standard yolov6m6 pretrained on the COCO dataset (I haven't trained it or any other model on the VisDrone dataset). So, which model/blob did you exactly mean? Best

Jan

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[tersekmatija]

Hey @drakshayanidesai, we don't do custom training for customers as part of our free support since we have limited resources and a lot of higher priority tasks. But we do offer several paid priority support packages. Feel free to reach out to support@luxonis.com if you are interested in that.

If not, we offer several training notebooks here which you can modify to train on VisDrone dataset. If GPU is running out of time in Colab, you could perhaps consider using Colab PRO, or download the tutorial and run it on any machine with dedicated GPU.

[drakshayanidesai]

Hi Jan

I trained model using VisDrone dataset but not able to detect Skatebords. Can you send u r model yolov6m6 which trained on COCO. Because in that it is possible.

Regdars Drakshayani

On Mon, 20 Mar 2023 14:08:03 +0530 Honza Čuhel wrote

Hi @drakshayanidesai, I am not sure if I understand you correctly, for the investigation described above I have used the standard yolov6m6 pretrained on the COCO dataset (I haven't trained it or any other model on the VisDrone dataset). So, which model/blob did you exactly mean? Best

Jan

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[tersekmatija]

Hey, @drakshayanidesai

You should be able to download yolov6m.pt from official release and use https://tools.luxonis.com/ to export it. Let me know if that works for you..

[drakshayanidesai]

Thank you very much

Using the model generated blob file. But unable to succeed. The model yolov6m.pt is pretrained on COCO dataset right?

Regdars Drakshayani

On Thu, 23 Mar 2023 22:58:17 +0530 Matija Teršek wrote

Hey, @drakshayanidesai You should be able to download yolov6m.pt from official release and use https://tools.luxonis.com/ to export it. Let me know if that works for you..

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[tersekmatija]

Yes correct.

[drakshayanidesai]

Let me check Image shape 640x320 right?

On Fri, 24 Mar 2023 19:36:00 +0530 Matija Teršek wrote

Yes correct.

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[tersekmatija]

Yes, or 640 x 352 should be closer to 16:9 aspect ratio.

[drakshayanidesai]

Hi

It is working fine for me. Thank you very much for your support. My target task is counting no of pedestrian and cyclist moving left/right. I know counting but directionality based not sure how to do.

Regards Drakshayni

On Fri, 24 Mar 2023 20:13:34 +0530 Matija Teršek wrote

Yes, or 640 x 352 should be closer to 16:9 aspect ratio.

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[tersekmatija]

You could use our object tracker node after Yolo to track the detected objects. Here's an example using a different type of object detector which you can modify. Then you need to look when object passes certain line/coordinates to increment the counter. We also have a fully working example here.

I am closing this issue since the main problem has been resolved. Feel free to open a new issue in the corresponding repository if the examples I linked above don't work for you :)

[drakshayanidesai]

Thank you very much your support

On Mon, 27 Mar 2023 15:57:06 +0530 Matija Teršek wrote

You could use our object tracker node after Yolo to track the detected objects. Here's an example using a different type of object detector which you can modify. Then you need to look when object passes certain line/coordinates to increment the counter. We also have a fully working example here. I am closing this issue since the main problem has been resolved. Feel free to open a new issue in the corresponding repository if the examples I linked above don't work for you :)

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[drakshayanidesai]

Hi

I need your guidance for "sunlight on trail" detection using Luxonis.

Regards Drakshayani

On Mon, 27 Mar 2023 15:57:06 +0530 Matija Teršek wrote

Closed #40 as completed.

—

Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID:

[tersekmatija]

Hey @drakshayanidesai , since it's not related to this issue, feel free to reach out to us on Discord, Discuss, or open a new GitHub issue in the appropriate repository.