A few months ago, we from PyImageSearch (where I work) took part in a CVPR competition. It concerns developing a model for accurate detection of natural scenes in mobiles. Today, I am glad to share that our entry made it to the top teams (good to know that we competed with ByteDance :D).
With respect to model size, efficiency, topline metric, etc. we think our model is decent enough. We have jotted down our solution approach in this report (contains approaches from other teams too): https://arxiv.org/abs/2105.08819. Our code is in TensorFlow and TensorFlowLite. We might be open-sourcing it to foster further research in the TinyML space.
Happy to address any feedback.