BoltNet accepted at CVPPA @ ECCV 2026

New Publication! Our paper “BoltNet: An Ultra-Lightweight Convolutional Network for On-Device Plant Species Identification” has been accepted at the CVPPA workshop (Computer Vision in Plant Phenotyping and Agriculture) at ECCV 2026.

BoltNet introduces a Spatial Redistribution Bottleneck and Logit PreSampling to tackle very large taxonomic spaces on constrained hardware. On Pl@ntNet-300K it reaches a 0.682 F1-score with only 341K parameters (1.37 MB), the best F1 among the evaluated models below 2 MB, and it was benchmarked end-to-end on CPU, GPU and NPU across Raspberry Pi 5, Jetson Orin Nano and Hailo-8.

Authors: Daniel Rossi, Guido Borghi, Roberto Vezzani Preprint: arXiv:2608.11844