YOLOv4-tiny custom - no detection about depthai-ml-training HOT 3 CLOSED

Hello @Whisper94 ,
I would like to apologize for the un-updated colab notebooks that have dead links and old openvino versions. We actually just updated all the notebooks in this PR #8. I believe Lukasz tested all the notebooks so they are working as expected and wrote instructions on how to run the blobs on the depthAI. You should check the PR branch for more info and follow those notebooks. I am sorry about the inconvenience.
Thanks, Erik

from depthai-ml-training.

Thank you Erik for fast answer.
I don't know what I did wrong but actual I'm confused because everything what I tried, every tutorial what I read has no success. Also default model tiny-yolo-v3 has NO detection but in console is another output. Yesterday I've tried it with blob converter but the same as earlier -> nothing happens. OpenVINO version is probably doesn't matter if they are match with blob file and script.

If I try my model with python depthai_demo.py -cnn custom_model_folder_under_nn --cnn_input_size 608x608 there is no results but this is output from console:

my converted blob files looks like "yolov4-tiny_custom_openvino_2021.3_5shave.blob". From tutorials I've changed JSON file, renamed it to like a blob file and put them in the same folder:

`{
"nn_config":
{
"output_format" : "detection",
"NN_family" : "YOLO",
"input_size": "608x608",
"NN_specific_metadata" :
{
"classes" : 2,
"coordinates" : 4,
"anchors" : [10,14, 23,27, 37,58, 81,82, 135,169, 344,319],
"anchor_masks" :
{
"side26" : [1,2,3],
"side13" : [3,4,5]
},
"iou_threshold" : 0.5,
"confidence_threshold" : 0.1
}
},
"mappings":
{
"labels":
[
"1",
"2"
]
}
}

`
however it doesn't work and I'm at the moment frustrated that simple things don't work as a should! Also the default version looks like with a bug or issue with my device?

with default parameters from tiny-yolo-v3 python depthai_demo.py -cnn tiny-yolo-v3 there is something with wrong CNN input size. But I don't know which HxW it was trained and where is it stored

I find ever different versions:
https://docs.luxonis.com/en/latest/pages/tutorials/first_steps/#using-custom-models
https://blog.roboflow.com/deploy-luxonis-oak/
https://colab.research.google.com/github/luxonis/depthai-ml-training/blob/use_blobconverter/colab-notebooks/Easy_TinyYOLOv4_Object_Detector_Training_on_Custom_Data.ipynb
https://colab.research.google.com/github/luxonis/depthai-ml-training/blob/master/colab-notebooks/Easy_TinyYOLOv4_Object_Detector_Training_on_Custom_Data.ipynb

Could you tell me what is wrong or provide me how can tell me it? Probably you need more information, let me know. At the moment I was only trying with default setting to be sure that everything works well but it isn't...

EDIT: I've changed config (json) file from default tiny-yolo with correct nn size 416x416 and it works. But my custom model can not detect anything although in colab version it works fine. I've renamed my json file accroding folder name
with python depthai_demo.py -cnn custom_model_folder_under_nn with blob file and custom_model_folder_under_nn.json there is no output (no detections, nothing in console). Now I don't know what is wrong. Probably 608x608 nn-size is to large for OAK-D WiFi although there are ~7-8FPS

from depthai-ml-training.

Hello @Whisper94,
I apologize for your dissatisfaction with the UX of the DepthAI. We are actively trying to improve it and resolve all the bugs as fast as possible. One example, as you mentioned, is the incorrect tiny-yolo-v3 size in the depthai_demo.py. We just merged the PR (luxonis/depthai#392) so it is fixed now.
For training, we advise you to use Roboflow, as it removed tons of complicated steps - steps where one can lose a lot of time, especially if it's the first time training. They even have an article about deploying the model to OAK-D, as you pointed out.
Our training examples are for users that want more customizability. And we have just updated all the notebooks with the PR I mentioned above.
If something isn't working as expected in the depthai repo, you can either open an issue on that repo or ask for help on our discord server - for an easier communication.
Thanks, Erik

from depthai-ml-training.