MobileViT-VO: Mobile Vision Transformer-based Visual odometry

Monocular visual odometry (VO) estimates six-degree-of-freedom camera motion from a single image stream and underpins autonomous navigation. Recent transformer-based VO models rival geometric pipelines in accuracy but need tens of millions of parameters and desktop-class GPUs, precluding deployment on the platforms that most need on-board odometry. We investigated whether a mobile vision transformer, designed for classification, can be adapted to continuous pose regression without sacrificing accuracy.

MobileViT-VO couples a MobileViT backbone with a spatio-temporal transformer. Three adaptations convert the classification architecture to regression: (i) correlation-aware motion tokens, concatenating consecutive feature maps with their difference and Hadamard product before temporal attention; (ii) a rotation head regressing a continuous nine-dimensional representation projected onto SO(3) by Gram–Schmidt orthogonalisation, supervised by a geodesic loss; and (iii) group normalisation with weight standardisation, stabilising regression at small batch sizes. Training used KITTI sequences 00, 02, 08 and 09 at 192x640 resolution, from scratch, with a single loss and optimiser; evaluation used held-out KITTI sequences 03-07 and 10 and the EuRoC MAV dataset, whose indoor aerial flight contrasts with KITTI's planar driving. Ablations isolated each adaptation.

MobileViT-VO requires 3.04M parameters and 1.71 GMACs per clip, 10.1x fewer parameters than TSformer-VO (30.66M). It achieved a 12% mean translational error and 8 deg/100m rotational error on KITTI, and a 3m absolute trajectory error on EuRoC. A mobile vision transformer can be adapted to monocular VO and attain accuracy comparable to substantially larger transformer models, provided motion is encoded explicitly, rotation is parameterised continuously, and normalisation suits small-batch regression. MobileViT-VO does so at an order of magnitude lower parameter cost, making transformer-based VO practical on resource-constrained devices. Monocular scale ambiguity remains a limitation.