我在edge impulse 训练的目标检测模型 未能成功在esp32-cam运行
但是我训练的分类模型Mobile net v1 ,可以成功的在esp32-cam 运行
Hi,
Object detection models are currently only supported on the platforms described in our docs page here: https://docs.edgeimpulse.com/docs/object-detection#1-prerequisites
We’re working on a new object detection pipeline for MCUs which will come out shortly, but not sure if it it will make it to ESP32 due to size / speed constraints.
@janjongboom can you please describe the issues with ESP32?
There should be no speed or size issues with latest examples repo from Espressif: https://github.com/espressif/tflite-micro-esp-examples
Hi @VikramDattu,
The examples linked in that repo are much simpler than object detection:
- hello_world - regression used to estimate a sinewave
- micro_speech - keyword spotting (classification)
- person_detection - image classification
Classification (including image classification) is much simpler than object detection. As a result, the object detection pipeline that Jan is referring to will require more RAM and CPU speed than those examples (and as such, may not run on an ESP32).
@shawn_edgeimpulse I understand.
With esp-nn
, which optimises ESP32 and ESP32S3 kernels, there is a significant speedup.
Let me know how it goes for your model and the bottlenecks you face.