Hi Everyone, I apologise if I've missed something obvious, I'm new to anything computer vision and ML related, and I'm struggling to follow the tutorial covered in [this Roboflow post](https://blog.roboflow.com/luxonis-oak-d-custom-model/), using [this Colab notebook](https://colab.research.google.com/drive/1fPn9WJQp_4dX1JfQNiBImcDwtEk_qa1b#scrollTo=gnNXNQCjdniL). My initial goal is to train the OAK-1 to identify good bananas (not over-ripe, not bruised) from bad-bananas (over-ripe, bruised, shrivelled). This is so I can get familiar with the process of annotating and training my own model, to expand to other cases. Steps taken so far: Annotated data sets in CVAT, 35 images for training and 13 for test. Run through the Colab tutorial, modifying the inputs to refer to my own data sets, rather than the ASL data sets. - My Colab notebook [here](https://colab.research.google.com/drive/1ePNnLGIiBFewRvntpx4voEjbQlVhZAY8?usp=sharing) - My data sets (exported as TFRecord data set from CVAT) [here](https://drive.google.com/file/d/16cN9zFQ-Ay0hEev-HgYUMtuahd2YhwTw/view?usp=sharing) and [here](https://drive.google.com/file/d/1VrKupdnXxLWUbzOCzDG3PMEKoGbcfssv/view?usp=sharing). - Everything in the tutorial seems to work OK (except installing Numpy in the first step?), but the output images don't seem to identify anything... it's just a bunch of boxes all over the image - see below: [URL=https://i.imgur.com/7NpoIfo.png][IMG]https://i.imgur.com/7NpoIfoh.png[/IMG][/URL] [URL=https://i.imgur.com/NEcN6Gb.png][IMG]https://i.imgur.com/NEcN6Gb.png[/IMG][/URL] I'm not totally sure what to try next - I have tried annotating the CVAT images with polygons (to better define the edges of the bananas) and also with rectangles (which is the data shared above) - I get similar results each time. I'd be very grateful for guidance on what to try next - if there's another guide I can follow, I'd be happy to give that a try also! Cheers, Andrew

Training a custom model with CVAT - Bananas - help needed!

S13_Mills

Hi Everyone,

I apologise if I've missed something obvious, I'm new to anything computer vision and ML related, and I'm struggling to follow the tutorial covered in this Roboflow post, using this Colab notebook.

My initial goal is to train the OAK-1 to identify good bananas (not over-ripe, not bruised) from bad-bananas (over-ripe, bruised, shrivelled). This is so I can get familiar with the process of annotating and training my own model, to expand to other cases.

Steps taken so far:
Annotated data sets in CVAT, 35 images for training and 13 for test.
Run through the Colab tutorial, modifying the inputs to refer to my own data sets, rather than the ASL data sets.
- My Colab notebook here
- My data sets (exported as TFRecord data set from CVAT) here and here.
- Everything in the tutorial seems to work OK (except installing Numpy in the first step?), but the output images don't seem to identify anything... it's just a bunch of boxes all over the image - see below:

I'm not totally sure what to try next - I have tried annotating the CVAT images with polygons (to better define the edges of the bananas) and also with rectangles (which is the data shared above) - I get similar results each time.

I'd be very grateful for guidance on what to try next - if there's another guide I can follow, I'd be happy to give that a try also!

Cheers,

Andrew

erik

Hello Andrew,
so the notebook you started with was from Roboflow but you removed the Roboflow part if I understand correctly? I personally don't know about this notebook (since it wasn't created by us) so I, unfortunately, can't help you with it, but we do have a set of our ml training demos here:
https://github.com/luxonis/depthai-ml-training/tree/master/colab-notebooks
You should give it a try (for example yolov4 or mobilenet), if you have any problem you can also try pinging us on Discord so it's less of a communication delay.
Thank you, Erik