Breaking01 / AI
Drone disaster model uses 0.1 MB of storage
The picoEmergencyNet researchers presented a design tailored to disaster classification instead of increasing model size.

Researchers said on 11 Nov that they had designed a disaster classification model requiring 0.1 MB of storage for use on drones.
The paper's abstract names the model picoEmergencyNet and lists 23k parameters and 7M FLOPs, a measure of computational work.
The researchers said the design extracts image features at different scales with little computation. It also combines channel attention, a mechanism that gives greater weight to important features.
The paper reported deployment experiments on NVIDIA Jetson AGX Orin hardware. The experiments ran without hardware-specific optimisations.
The researchers said the high computational cost of existing approaches limits their deployment on drones with constrained resources. They presented an architecture tailored to the task instead of increasing model size.
The researchers did not disclose disaster classification performance during actual flight in the abstract.