Using the ESP-Detection framework for on-device insect detection in images acquired by an embedded vision system.

[Jasppt]

Checklist

• Checked the issue tracker for similar issues to ensure this is not a duplicate.
• Provided a clear description of your suggestion.
• Included any relevant context or examples.

Issue or Suggestion Description

The ESP-Detection framework was used to train an object detection model on a dataset of ~2,500 labeled images of a target fly species. Annotations were created with LabelImg and converted to YOLO format. The model was trained at input resolutions of 640×640, 416×416, and 224×224. When deployed on the ESP32-P4, the model does not detect the target insect during inference, even with high-quality input images.

[Ginosko-mia]

Hi, @Jasppt
Can the float model detect your target?

[Jasppt]

Yes . I trained same dataset in colab google and tested with good results.

[Ginosko-mia]

@Jasppt As you mentioned, the espdet_pico float32 model works properly, but the quantized model fails to detect the target insect on the ESP32-P4. This suggests that there may be a significant accuracy drop after quantization.

Could you please upload your trained .pt model? I would like to use it to help debug and analyze the quantization process for your model.

[Jasppt]

weights.zip
Good morning.
Send in atachment diferent train result in diferent resolutions.

Regards

[Ginosko-mia]

@Jasppt I need some validation images as well, like 20 images. Can you upload some?

[Jasppt]

calib.zip

Sure.
Send in atachment.
Regards

[Ginosko-mia]

Hi, @Jasppt
Can you show me your good test results with the float32 model? Since I test the float32 model first and found that only train4 can successfully detect the target. But the conf and bbox is not ideal.

As a result, the model easily fails after quantization (default confidence threshold = 0.25). Therefore, I believe this issue may originate from the model itself rather than the quantization process.

Moreover, I found the training set is not ideal. Some wrong data exist.