ROMFS Support is here!

Posted by Kwabena Agyeman on June 04, 2025

Hi Everyone!

It's finally here! All OpenMV Cams now support ROMFS with the release of OpenMV firmware v4.7.0 and OpenMV IDE v4.7.0!

What is ROMFS?

Glad you asked! So, we've always embedded large binary files in our firmware like Haar Cascades and Neural Networks. As firmware space has gotten tight over the years as we've added new features, fixed bugs, and so on and so forth we've had to enable/disable assets across different cameras to make space.

We know that this has caused a lot of frustration.

OpenMV Cams have always had a FAT file system onboard and SD card support. However, any files read from FAT have to be loaded into RAM... which has also always been limited.

Enter ROMFS

ROMFS solves the storage problem by creating a memory-mapped file-system which you edit via OpenMV IDE! All files stored in the new ROMFS partition on your OpenMV Cam can be read like as if they were files, but, we don't have to load them into RAM. This feature allows you to deploy large binary assets like Neural Networks to your OpenMV Cam, which can be executed in-place off of the flash.

External flash put to use

Now with ROMFS, we've put the external FLASH on the OpenMV Cam H7 Plus, RT1062, Arduino Portenta, Arduino Nicla, and Arduino Giga to good use. ROMFS allows us to offer MB's of storage now that can be used for larger ML models without requiring any RAM.

Furthermore, on the upcoming OpenMV AE3 and N6 we've really leaned into leveraging ROMFS with 24MB partitions for model storage. Now you can run 10MB+ YOLOV8 models on our new cameras easily!

Under the hood changes

ROM support required quite a few monumental changes:

To support reading and writing ROMFS, we updated all our OpenMV Cam bootloaders (except the RT1062) to use DFU. Now, with the latest bootloaders you can load your OpenMV Cam's firmware using standard command line tools (which OpenMV IDE ships with). For the RT1062, you have to use NXP's tools, but, it has a working ROMFS too!
OpenMV IDE had to be upgraded to handle bootloading OpenMV Cams with the old bootloader and new bootloader at the same time, so folks can seamlessly upgrade.
MicroPython had to support ROMFS upstream after three years of development.
We had to update our firmware and automated build system to move all large binary assets to the new ROMFS partition for each camera.
OpenMV IDE had to be updated to support reading and writing ROMFS in all firmware update interactions.

This took a long time! But, it's very much worth it! Now, with ROMFS, we can now leverage large custom binary assets without any issues.

ML models? No problem. Giant remap tables for lens correction? Coming soon.

Other Interesting Technical Tibdits

The ROMFS file system stores file records in memory, aligned, contiguous chunks in flash. The contiguous part is why they don't have to be read into RAM, whereas FAT stores files in chunks, which can be all over the place. ROMFS also allows us to memory align files to 16-byte, 32-byte, and etc. addresses. This superpower allows us to pass the address of a neural network in ROMFS directly to the NPU hardware accelerators onboard the OpenMV AE3 and OpenMV N6.

Now Check it Out!

This is what ROMFS allows us to do:

Now we can run big models onboard OpenMV Cams!

And the New Model Zoo!

Last but not least, we now integrate a Model Zoo! By the time we release the OpenMV AE3 and N6, we're going to have demo model support for:

Monocular Depth Estimation from the Camera Image
Face Detection
Facial Landmark Detection
Facial Recognition
YOLOV2, YOLOV5, YOLOV8, and more objection detection
Pose Estimation
Instance and Semantic Segmentation
And more!

You can check out the models we plan to support via Tools->Machine Vision->Open Model Zoo.

Model Conversion is built in too!

Not only does OpenMV IDE ship with the Model Zoo, but it also integrates ARM's Vela Compiler to convert Models to run on the OpenMV AE3 and ST's EdgeAI tools to convert Models to run on the OpenMV N6! Now you can easily bring your own INT8 quantized TensorFlow Lite Model and run it onboard with NPU acceleration.

Roboflow Integration!

And as promised during our Kickstarter, Roboflow is working on enabling YOLOV8 INT8 quantized TensorFlow Lite Model output support, so you can train your own custom models too!

And MPY Cross!

Last but not least, OpenMV IDE also includes MPY Cross too so you can compile your Python code to bytecode to store on ROMFS or the FATFS. We just support compiling individual scripts now, but, we'll add support for compiling modules next.

Kickstarter Progress

We're steadily working towards delivering the OpenMV AE3 and N6 this year. Here's where we are so far:

OpenMV IDE: On the desktop tool side, the OpenMV IDE is feature-complete to support the OpenMV AE3 and N6. Not only has it been updated to support ROMFS, the new Model Zoo, and all the new bootloaders, but, its also ready to fully provision and recovery the OpenMV AE3 and N6 using vendor tools so we can manufacture the OpenMV AE3 and N6 easily along with also being able to always factory reset your OpenMV Cam.

OpenMV AE3: We're sending out our final prototype for the OpenMV AE3 this week, along with ordering long-lead time parts for mass production. From the looks of things, we're on track to start delivering the OpenMV AE3 by the end of Q3 this year. Everything is pretty much operational onboard, we have some camera image quality tuning to do still, but otherwise, we just need to get all our demo models from the model zoo working onboard with example scripts.

OpenMV N6: The OpenMV N6 MicroPython port is still in progress. However, we've pretty much verified that all the hardware on the board is working as expected. For the new multispectral camera support, we now have a new camera system interface which lets you use multiple frame buffers at the same time!