FOMO receives objects for esp32

#1

Hello. I am interested in a project to identify motorbikes on the road from surveillance cameras. But when teaching the model, we see that the model misidentifies quite a lot. So can I ask how to identify more accurately? Or show me how to get up for the best results? I use esp32 s3. Project ID: 404351

#2

Hi @Tien,

Object detection can often be quite tricky. I recommend trying a few things to improve accuracy:

  1. Get more data. Many object detection models (even with transfer learning) require hundreds or thousands of samples to get a reasonable model working.
  2. Train for longer. Object detection training often requires hundreds or thousands of epochs.
  3. Try larger input images. It looks like your images are scaled to 96x96, which is quite small. I recommend trying higher resolution (e.g. 240x240) to start.
1 Like
#3

I tried with higher resolution but when running the model it will take more RAM. I’m worried that the ESP 32’s ram won’t be enough

#4

Hi @Tien,

If you have access to EON Tuner, I also highly recommend trying that to maybe find a model architecture will work.

#5

I have access to EON Tuner but the models all output with 96x96 resolution. And with current image data, this is the model with the most accurate results compared to other models. I also received more data but found that the accuracy did not improve and even decreased. We look forward to receiving solutions and help from your team.

#6

Hi @Tien,

First, it looks like your data has varying resolutions (e.g. from 60x60 to 262x262). If you’re able to get higher resolution data, that might help. Second, on the “Create impulse” page, you can set the input image resolution (all images will be scaled/cropped to this resolution). You might try increasing this to something like 240x240 to see if that helps.

1 Like