EVEE

Core Idea

Events become online supervision

Offline-trained local feature detectors and matchers often degrade on unknown targets because training data cannot cover every future object appearance and structural variation. EVEE uses event streams as proxy supervision to adapt a lightweight keypoint head online, enabling label-free target-specific feature learning during deployment.

The framework combines Training Warehouse for fast adaptation, the temporal voting bank for foreground and background memory, and ReWeight for reliability-aware correspondence filtering. Together, they improve unknown-target matching and pose estimation while keeping real-time matching speed.

Event-guided adaptation Events reveal which RGB regions should become learnable features.

Looping explainer

01RGB frame

event stream

Events traverse the appearance

Event-guided
selection

02Activated RGB

Reliable regions become features

Online
update

03Feature learning

Online head

Target-specific representation

Method

EVB

Event Voting Bank

EVB keeps event-derived foreground and background voting memory. Consistent target activity is reinforced over time, while short-lived background noise is forgotten before matching.

TraW

Training Warehouse

TraW runs multiple online training threads in parallel. The thread with the lowest loss becomes the current best weight, then all threads continue from that stronger checkpoint for the next update.

ReWeight is kept as a compact matching gate after adaptation, but the project page focuses on EVB and TraW because they explain the main online behavior of EVEE.

How TraW works

Parallel trial, shared best weight.

Live explainer

Results in motion

Browse TUD-L, NAVI-MultiView, and three public real event-camera sequences. Each visualization stays paired with its own quantitative comparison.

05 experiment views
one synchronized story

68.7 Average result on NAVI and TUD-L

51.3 FPS Online matching speed

14.52 FPS Online adaptation speed

+4.4% Gain over the second-best method on public event-camera data

Temporal match stream

Tracks persist across changing illumination and target scale in a continuous household-object sequence.

TUD-L

Sequential unknown targets

RGB-D benchmark

Scale~24K frames SignalRGB + simulated events MetricPose ACC / AUC

Method	ACC			AUC
Method	5°	10°	20°	5°	10°	20°
SP + SuperGlue	72.7	75.4	78.8	69.9	72.0	74.6
SIFT + LightGlue	70.7	75.7	82.9	67.2	70.2	74.9
DISK + LightGlue	71.8	75.0	79.5	68.5	71.0	74.2
DoGHardNet + LG	70.3	75.5	82.7	66.8	69.9	74.7
ALIKED + LightGlue	71.5	76.0	82.5	67.9	70.9	75.2
SP + LightGlue	72.0	76.1	81.7	68.7	71.4	75.2
OmniGlue	74.5	77.0	80.2	71.5	73.7	76.2
EVEE	77.0	78.9	81.0	74.8	76.4	78.3

Bunny racer sequence

Dense, coherent correspondences remain visible around the head, ears, and other thin object structures.

NAVI-MultiView

Cross-view object matching

Unseen targets

Scale~25K image pairs SignalRGB + simulated events MetricPose ACC / AUC

Method	ACC			AUC
Method	5°	10°	20°	5°	10°	20°
SP + SuperGlue	54.3	79.3	87.2	20.9	45.9	65.0
SIFT + LightGlue	52.9	76.4	87.0	21.8	45.0	63.8
DISK + LightGlue	46.1	70.4	81.0	18.1	39.9	58.3
DoGHardNet + LG	53.2	76.5	87.2	22.0	45.2	64.0
ALIKED + LightGlue	50.3	73.8	84.2	20.8	43.1	61.6
SP + LightGlue	55.1	78.4	88.3	22.5	46.4	65.3
OmniGlue	49.2	73.8	81.7	18.6	42.1	60.4
EVEE	54.6	81.4	89.5	20.3	46.4	66.3

Fixed response ROI

Temporal tracks rendered directly from the local shapes-6DoF RGB, event, and EVEE match outputs.

Shapes 6-DoF

Real event-camera motion

Seen reference

Motion6-DoFSignalRGB + real eventsRolePre-training reference

Method	ACC			AUC
Method	5°	10°	20°	5°	10°	20°
SP + SuperGlue	54.8	62.4	68.1	40.3	49.6	57.7
SIFT + LightGlue	55.7	60.8	64.8	41.1	50.0	56.6
DISK + LightGlue	63.4	70.2	74.0	50.7	59.0	65.8
DoGHardNet + LG	54.1	60.3	64.7	40.1	48.8	55.9
ALIKED + LightGlue	41.8	47.1	49.6	30.4	37.7	43.2
SP + LightGlue	56.1	61.1	65.2	41.6	50.3	57.0
OmniGlue	55.5	63.3	68.5	39.7	49.7	58.1
EVEE	65.5	73.5	77.1	47.2	58.9	67.3

Fixed response ROI

Fast camera motion and background event noise test whether target-specific correspondences remain stable.

Boxes 6-DoF

Unseen event-camera sequence

Unseen sequence

Motion6-DoFSignalRGB + real eventsSplitUnseen

Method	ACC			AUC
Method	5°	10°	20°	5°	10°	20°
SP + SuperGlue	66.7	75.2	76.9	46.7	59.4	67.9
SIFT + LightGlue	62.7	71.5	72.5	44.0	56.1	64.2
DISK + LightGlue	70.5	78.2	79.8	51.2	63.2	71.3
DoGHardNet + LG	65.0	72.3	74.0	45.7	57.6	65.5
ALIKED + LightGlue	63.0	70.0	71.5	44.2	55.7	63.4
SP + LightGlue	71.1	77.6	78.9	51.4	63.3	71.0
OmniGlue	65.4	72.7	75.0	45.7	57.9	66.1
EVEE	71.1	80.6	85.4	49.7	63.3	73.5

Fixed response ROI

The local poster-6DoF outputs show feature trajectories over textured content during free camera motion.

Poster 6-DoF

Unseen event-camera sequence

Unseen sequence

Motion6-DoFSignalRGB + real eventsSplitUnseen

Method	ACC			AUC
Method	5°	10°	20°	5°	10°	20°
SP + SuperGlue	51.1	58.6	65.0	36.6	46.2	54.1
SIFT + LightGlue	54.5	62.2	68.3	39.1	48.9	57.0
DISK + LightGlue	53.7	60.4	66.0	37.7	47.7	55.4
DoGHardNet + LG	53.1	60.4	66.7	38.5	48.0	55.8
ALIKED + LightGlue	52.6	58.9	64.8	37.4	46.8	54.6
SP + LightGlue	53.4	60.3	66.6	38.4	47.8	55.7
OmniGlue	53.1	60.0	65.5	38.0	47.6	55.3
EVEE	62.4	75.6	79.9	42.0	56.2	67.3

Acknowledgements

This work was supported by Japan Science and Technology Agency Support for Pioneering Research Initiated by the Next Generation (JST SPRING) under Grant Number JPMJSP2109.

Citation

Cite EVEE

@inproceedings{evee2026,
  title     = {EVEE: Event-Based Online Adaptation for Matching on Unknown Targets},
  author    = {Zhao, Zejing and Ju, Cheng and Zhang, Yanwen and Namiki, Akio},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}

Events become online supervision

Overall EVEE structure

Event Voting Bank

Training Warehouse

Parallel trial, shared best weight.

Sequential unknown targets

Acknowledgements

Cite EVEE