Trying to deploy my custom object detection model using esp32cam

Projects
1

im trying to delpoy my custom object detection model with (2 classes) on ESP32CAM but i facing so many errors, and couldnt if there is any tutorial or video than can help me. any help would be great.

2

I know that @mjrovai has had success with the esp32cam before. For example, ESP32-CAM: TinyML Image Classification - Fruits vs Veggies - Hackster.io

3

i have seen this before but he is using edge impulse im using yolov5 and colab cuz when i tried edge impulse the model map was so low i dont know why. plus its object detection.

4

Ah gotcha well unfortunately I don’t have much experience with the ESP32s but hopefully someone else can jump in and help!

1 Like
5

Hi, @saeedbenadeeb101 and @brian_plancher, I also never tried Object Detection on ESP32CAM, but I am not sure that the device has enough memory to run a Yolo5 model on it. How much is the size of your model? That’s why Edge Impulse comes with FOMO that, in the end, uses a modified Image Classification model (smaller model) as a basis for object detection.

6

Hi, @mjrovai. My model tflite model size is 400kb and when i convert to .h model its 2.69mb

7

Hi. The ESPCAM has only 520KB of SRAM, which should be the main problem besides the CPU’s low capability. If you need Object detection running fully on a Tiny device, I suggest the Seeed Grove Vision AI Module (Getting Started - Seeed Wiki). It is a $26.00 module with a Himax HX6537-A processor, thumb-size AI-powered OV2640 camera sensor, supporting YoloV5. I tested a project using Roboflow (https://roboflow.com) to annotate my dataset and create a YoloV5 object detection model. The inference was done entirely on the device and worked well. You can follow the tutorial Vision AI Train Own Model - Seeed Wiki.

1 Like
8

Hi, I appreciate your suggestion to run on Seeed Grove Vision module, but the i noticed that on ESP32CAM library there is like a face detection model, but the resolution needs to be less than 320x240, that’s why i trained my model on 240x240 so i can deploy my model on ESP32CAM, but when i tried to deploy i didn’t find any guide on how to do so. plus, do u think if i can reduce the file size more than that i can deploy my model, if yes what should the size to run on ESP32CAM. the module u suggested is not available on Malaysia and delivery from overseas could take more than 6 weeks that’s why I’m trying to run on ESPCAM.