Custom model using SDD scene

[DevDigitalCittius]

Hi

I'm having some troubles using my custom TFlite model, in the SDD scene.
I trained my model using this google collab notebook , and the result is kind of acceptable for the testing that im doing, as is shown in the image below.

When i try to use it in the SDD scene, i get this error message.

The only way i found to use my model in the SDD scene, is quantizing the model, when i quantize it, the model runs normally in the SDD scene, but there is no detection at all (look at image bellow)

After a lot of trial and error, i found out, that when i quantize my model, it just stops working as expected.
Is there a way I can use my model without receiving this error message? Or a way I can quantize it without crashing my model?

Here is the tflite file for the working model

Sorry if i got something wrong, it's my 1st time training a tensorflow model

Thanks in advance

[asus4]

Hi @DevDigitalCittius,

Change BaseImagePredictor<sbyte> to BaseImagePredictor<float>. This allows the use of non-quantized 32-bit float input.

tf-lite-unity-sample/Assets/Samples/SSD/SSD.cs

Lines 9 to 10 in b7692d1

	public class SSD : BaseImagePredictor<sbyte>
	{

Then, the output order is a bit different from this old example. change the outputs like the following ↓
Please check out #312 (comment) for more details.

Screen shot:

[DevDigitalCittius]

Hi @asus4
I made those changes and it's working!!
I think I just need to train my model a bit more.

Thanks for the quick repply
Best regards.

[asus4]

Good luck 👍

I plan to update this Object Detection example from SSD MobileNet v2 to a more recent model—maybe EfficiantDet Lite, which might be easier to train the custom model.

https://developers.google.com/mediapipe/solutions/vision/object_detector