The KITTI Vision Benchmark Suite

PAVO is a lightweight self-supervised monocular
visual odometry method that predicts relative
camera motion from consecutive left color images.
The method builds on a SC-Depth-style depth-pose
framework and introduces point-aware pose learning
through learned matching confidence, motion-
descriptor residual guidance, and a rotation-
decoupled ARN branch. During inference, PAVO
performs single-pass pose prediction without
stereo input, LiDAR, loop closure, bundle
adjustment, or factor graph optimization. The
submitted trajectories are generated by
sequentially composing the predicted frame-to-
frame poses with the same model parameters for all
KITTI test sequences.

Input: monocular left color images only.
Training data: KITTI odometry sequences 00–08.
Test data: KITTI odometry sequences 11–21.
Image resolution: 640 x 192.
Backbone: ConvNeXtV2-Atto with 1/8-resolution pose
feature stage.
Pose estimation: direct frame-to-frame regression,
no test-time BA, no factor graph, no loop closure.
Matching guidance: confidence residual + gated
motion-descriptor residual.
ARN branch: rotation-decoupled auxiliary rotation
head.
Depth range: 0.1–100 m.
Same parameters are used for all submitted
sequences.

@misc{anonymous2026pavo,
title = {Point-Aware Self-Supervised
Visual Odometry with Rotation-Decoupled Pose
Learning},
author = {Anonymous},
year = {2026},
note = {Manuscript under double-blind
review}
}