The training performance rate is very low when I use the FOMO algorithm

#1

Hi , My project ID is 14295255
My project is object detection by using Arduino Nano 33 BLE Sense .My problem when I use the FOMO algorithm the train performance is too less also model testing.
For information, when I use the SSD algorithm, I get 89% training performance but when i use FOMO i get only 12% !!

1 Like
#2

Welcome @eng771.ruqaya.alaa

It’s not uncommon to see different performance rates between algorithms like SSD and FOMO. Particularly in smaller sized datasets.

This can be improved by having better quality datasets. SSD it is worth noting will not give you the same range of platform support, or speed of operation, and the BLE33 in particular will struggle. You may need to try alternative hardware if accuracy on larger sets of classes is key over speed and volume of objects. There is no golden model, and you need to choose the correct tool for the task.

The FOMO (Faster Objects, More Objects) algorithm, is an approach for real-time object detection on microcontrollers. It’s significantly faster (30x than MobileNet SSD) and requires less RAM (<200K). FOMO is designed for situations where object locations are more critical than their sizes, using a centroid-based detection method instead of traditional bounding boxes. It works well with objects of similar size and should not be too close to each other. FOMO’s reduced complexity and size make it suitable for constrained devices but may result in lower accuracy compared to more complex models like SSD, explaining the difference in performance you observed in your project. For detailed usage and optimization, you can refer to the FOMO documentation.

To improve your FOMO model’s performance, consider the following:

  1. Data Quality: Ensure your dataset is well-labeled and diverse enough to cover the variations in objects you want to detect.
  2. Model Parameters: Experiment with different hyperparameters, learning rate etc.
  3. Augmentation: Data augmentation (like rotating, flipping, or adding noise) can help the model generalize better.
  4. Simplifying the Task: If possible, simplify the object detection task. FOMO should perform better with fewer classes or more distinct objects.
  5. Distance from the object FOMO works best if the distance from the object is constant across training testing and usage. See the docs below for a more in depth guide.

Hope this helps, again if you do have some further analysis to share please do, I will try to answer or someone else will be able to review your findings.

Best

Eoin

1 Like
#3

Thank you very much f or your reply @Eoin:pray: