Rose Flower Classification with OAK-D

FabianRamirez

I work on a rose (flower) classification project, I have trained a neural network to identify the flower that is ready for cutting and the one that is not for cutting using OAK-D. I have to do a tracking following the example of the object tracker on RGB tutorial, it is it makes use of the MobilNet pre-trained network.. but I can't get it to work.. I'm using my own .blob file that I generated with TensorFlow in Google Colab... can someone give me a hand on this issue thanks.

erik

Hi FabianRamirez ,
I think you'd want to train custom object detection model for roses, instead of classification, and use that together with ObjectDetection / ObjectTracker node, so you can do both detections and tracking of rose flowers. Thoughts on such an approach?
Thanks, Erik

FabianRamirez

Hi Erik, thank you for your answer, that is to say that I should train a model only for rose detection, right?... the model that I have trained is for classification, therefore it is not useful for me, is it?

I will try what you say, thank you very much.

Later, I must achieve a rose count, the project is heavy hehehe, I hope I can achieve it.

jakaskerl

Hi FabianRamirez,
Why do you need the tracking? I assume you intend to tell/infer for each rose flower whether it's ready for cutting and probably return a number of roses that have to be cut. For this you would need a model that is great at detecting separate flower and than passing each flower to the already created classifier.
I'm not sure where the object tracking comes in.

Thanks,
Jaka

FabianRamirez

Hello Jakas, thank you very much for answering, I must do a count too, in the example models that I have seen they use tracking, that is why I think I should do a tracking to count, I am open to opinions, I have not been on this topic for a long time, thank you.

jakaskerl

Hi FabianRamirez

Do you intend to run the camera trough a line of flowers and decide how many need cutting? Maybe you could take an image every # seconds and count the flowers that need cutting inside all images combined. It really depends on the use-case.

Thoughts?

FabianRamirez

Hello Jakas, yes that is exactly what I want to do, but first I want to train a network that can count all the roses, from the ones that are growing to the most developed ones, I have designed a device that is transported by a cable like a cable car and carries the camera on board, take video above the beds or rose crops... I'm following the DepthAI tutorials but I still have doubts on how to use tracking or object counting... but mainly I have to count all the roses in cultivation

jakaskerl

Hi FabianRamirez
In order for your Neural network to work with object tracker, you need it to output the image detections message, so object tracker can read it. This would mean that the model is capable of detecting a flower ready for cutting and drawing a bounding box around it. Is your current model capable of doing that?

Thanks,
Jaka

FabianRamirez

At the moment it doesn't, that's what I want to achieve, but I still have the question if I should train a network with the images of many roses and then do the tracking... In short, I plan to do the following:

train the neural network with images of crop rose where I will apply the pilot project in classification mode to generate my .blob file

to do the tracking and based on this then do a count, for the moment it doesn't matter if the flower is cut, I want to be able to count all the roses that the camera sees..

thanks

jakaskerl

Hi FabianRamirez
If you don't need to know if the flower is cut or not, you will need a detection NN that will find all flowers in a given image. Then you can easily apply the object tracking/counting like in the people counting example.
To start you can checkout the training guide.

Hope this helps,
Jaka

FabianRamirez

Thanks, I'm going to try this... later, if I have to implement a code that allows me to know which flower is to be cut according to its state of development... thank you very much Jakas