TLEIO: Tight Learned Event-Inertial Odometry

Advances in Event-Based Computer Vision @ ACCV 2026

12*Equal contribution

TLEIO estimates 6-DoF motion from a raw event stream and an IMU, without frames, feature tracks or pose optimization.

Abstract

Event cameras provide microsecond temporal resolution, high dynamic range, and sparse asynchronous measurements, making them compelling for odometry under fast motion and challenging illumination. Yet, accurate event-based state estimation remains difficult: existing systems often use events as a complement to frames or rely on feature tracking, correspondence search, and optimization-heavy backends, limiting the efficiency promised by event sensing. We introduce Tight Learned Event-Inertial Odometry (TLEIO), a compact filter-based pipeline that learns short-window relative motion directly from voxelized event streams and tightly fuses the resulting constraints with IMU propagation. Its learned front-end, EventsFormer, predicts local displacements and per-axis uncertainty from event voxels in a single forward pass, while a stochastic-cloning EKF fuses these overlapping predictions with IMU propagation, producing metric trajectories without explicit feature tracks and pose optimization at inference time. The network is trained directly from pose displacements, without requiring event-flow labels. We evaluate TLEIO on the TartanAir/TartanEvent benchmark and on real DAVIS240c sequences. On TartanAir, TLEIO reduces average ATE from 2.26 m to 1.10 m over the IMU-only TLIO baseline, while achieving up to 48.9% lower runtime than optimization-based event odometry methods.

Results

We ran the released checkpoint end to end on the fifteen sequences of the TartanEvent office environment. Pick a sequence to see the raw events next to the estimated 3D trajectory and orientation. The estimate is drawn exactly as the filter produces it: it starts at the true pose, and any gap to the dashed ground truth is accumulated drift.

Easy
Hard

The office environment is part of the TartanEvent training data, so these sequences illustrate the pipeline rather than held-out accuracy (see Benchmark below). The 200 Hz IMU is derived from ground truth, and the filter starts from the ground-truth pose. Two filter parameters, the network displacement scale and the measurement-covariance scale, were calibrated once for all fifteen sequences together (0.85 and 4.0), never per sequence; the four IMU noise densities were searched as well but changed the result by less than 0.01%. The videos show the unaligned estimate. For ATE, the estimate is rigidly SE(3)-aligned to ground truth over the whole sequence (scale fixed); across the fifteen runs the median ATE is 1.55 m, or 1.7% of the path. Videos play in real time.

Method

TLEIO splits odometry into a learned front-end that reads motion from events and a lightweight filter that integrates it with inertial data.

TLEIO overview: event camera, preprocessing, tokenizer, EventsFormer and prediction head, fused with IMU propagation in a stochastic-cloning EKF
  1. Event voxels

    Events are accumulated in 50 ms windows of five temporal bins.

  2. EventsFormer

    A clip of consecutive voxels is split into patches and processed by a transformer with divided space–time attention. It regresses each short-window displacement together with a per-axis uncertainty, in a single forward pass.

  3. Stochastic-cloning EKF

    IMU measurements propagate the state at high rate. Each learned displacement becomes a relative-position update between cloned past poses, weighted by its predicted uncertainty. No feature tracks and no pose optimization.

EventsFormer architecture

Tokenizer: each event voxel is split into patches, projected to tokens, with a temporal embedding
Tokenizer. Each event voxel is split into patches and projected into tokens, with a temporal embedding for the voxel index.
EventsFormer encoder: temporal then spatial attention with RoPE2D, repeated
Encoder. Temporal and spatial attention are applied in sequence, with RoPE2D in spatial attention.

Benchmark

Results reported in the paper, on the held-out TartanAir competition split, on real DAVIS240c recordings, and for end-to-end runtime on EDS.

1.10 maverage ATE on TartanAir,
vs. 2.26 m for IMU-only TLIO
0.45 %average error on real
DAVIS240c sequences
48.9 %up to this much lower runtime than
optimization-based event odometry
ATE per TartanAir competition sequence for TLIO and TLEIO
TartanAir competition split, ATE after SE(3) alignment. Adding learned event constraints to the IMU-only filter lowers the error on 12 of 16 sequences and halves it on average.
End-to-end runtime of TLEIO, RAMP-VO, DEVO and DEIO on EDS 00, 01 and 09
End-to-end runtime on EDS, from raw events to trajectory, mean of five runs. TLEIO is fastest on two of three recordings and overall; the filter accounts for only 5.4% of its runtime.
Table 1: TartanAir competition split. ATE [m]; lower is better. Published baselines are reproduced from the RAMP-VO benchmark. F, E, and I denote frames, events, and IMU; bold marks the lowest error in each row.
Seq.EKLT-VIOORB-SLAM3COLMAPDROID-VODPVORAMP-VOTLIOTLEIO
InputE+F+IFFFFE+FIE+I
ME0012.8013.6115.200.220.160.201.452.13
ME019.1016.865.580.150.110.042.301.20
ME0216.2320.5710.860.240.110.101.421.45
ME0318.0816.003.931.270.660.467.311.37
ME0416.9622.272.621.040.310.161.141.37
ME056.849.2814.780.140.140.131.851.19
ME0621.4421.617.001.320.300.122.470.48
ME079.067.7418.470.770.130.121.370.72
MH0014.6815.4412.260.320.210.362.392.65
MH014.352.9213.450.130.040.060.940.14
MH0210.2013.5113.450.080.040.041.611.16
MH032.138.1820.950.090.080.043.720.22
MH042.722.5924.971.520.580.411.870.60
MH054.4821.9116.790.690.170.251.350.48
MH066.9511.707.010.390.110.111.251.15
MH0719.7225.887.970.970.150.071.661.44
Average10.9814.3812.500.580.210.172.261.10
Table 2: DAVIS240c held-out sequences. MPE [%]; lower is better. Published baselines are reproduced from the DEIO benchmark. F, E, and I denote frames, events, and IMU; bold marks the lowest error in each column.
MethodInputboxes trans.HDR boxesboxes 6DoFposter trans.HDR posterposter 6DoFAvg.
EKLT-VIOE+F+I0.480.460.840.350.650.350.52
Lee et al.E+F+I0.740.690.770.280.520.590.60
Rebecq et al.E+I0.570.920.690.890.590.820.75
HASTE-VIOE+I2.551.752.031.340.571.501.62
Zhu et al.E+I2.691.233.610.942.633.562.44
Ultimate-SLAME+I0.760.670.440.150.490.300.47
DEIOE+I0.070.090.050.040.060.080.07
DEVOE0.060.060.710.060.140.440.25
TLEIOE+I0.310.550.490.430.500.430.45

BibTeX

@misc{cascone2026tleio,
  title         = {Tight Learned Event-Inertial Odometry ({TLEIO}): a Transformer-Based
                   Architecture for End-to-End Motion from Event Streams},
  author        = {Cascone, Matteo and Jensen, David and Pirini, Alessandro and
                   Stefanini, Angelo and Ven{\`e}, Pietro and Tintor{\'e} Gazulla, Marta and
                   Hidalgo Carri{\`o}, Javier},
  year          = {2026},
  eprint        = {2610.33356},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}