Eva-Tracker: Multi-Modal Event Tracking

Updated 5 July 2026

The paper on event-camera feature tracking introduces a deep, data-driven pipeline using grayscale templates and event patches, achieving up to 130% improvement in relative feature age with low-latency processing.
The aerial tracking framework employs an ESDF-update-free, visibility-aware planner to generate collision-free trajectories and target reacquisition paths, demonstrating low occlusion and failure rates in real-time simulations.
GS-EVT, the cross-modal approach, aligns integrated event images with Gaussian Splatting renderings to optimize camera pose and velocity, delivering competitive accuracy in absolute trajectory error compared to conventional methods.

Eva-Tracker is a name that, in the literature represented here, does not denote a single canonical method. It refers directly to a deep event-camera feature tracker introduced in “Data-driven Feature Tracking for Event Cameras” (Messikommer et al., 2022) and to “Eva-Tracker: ESDF-update-free, Visibility-aware Planning with Target Reacquisition for Robust Aerial Tracking” (Lin et al., 13 Feb 2026). In query-level usage, the label is also applied to GS-EVT, a cross-modal 6-DoF event-camera tracker that aligns integrated event images with differential renderings from a Gaussian Splatting map (Liu et al., 2024). The common thread is tracking under regimes where conventional frame-only pipelines or repeated ESDF updates are limiting; the tracked variable, sensing stack, and optimization strategy, however, differ substantially.

1. Nomenclature and scope

Within the provided literature, “Eva-Tracker” spans point-feature tracking, aerial target tracking, and event-camera pose tracking. The term is therefore best read as a disambiguation problem rather than as a single standardized acronym.

Usage	Domain	Distinguishing mechanism
Eva-Tracker (Messikommer et al., 2022)	Event-camera feature tracking	Grayscale template patch, event patches, frame attention module
Eva-Tracker (Lin et al., 13 Feb 2026)	Aerial target tracking	FoV-ESDF, visibility-aware initial path generation, target reacquisition
GS-EVT as related query usage (Liu et al., 2024)	Cross-modal 6-DoF event-camera tracking	Gaussian Splatting map, local differential rendering, coarse-to-fine optimization

The first of these systems tracks point features detected in a grayscale frame through subsequent event streams. The second is a trajectory-planning framework for a drone that must keep a moving target visible while avoiding occlusions and collisions. The third estimates the pose of an event camera by comparing integrated events with rendered intensity changes from a frame-built 3D Gaussian Splatting scene representation. This scope difference is foundational: the state being estimated is, respectively, a feature displacement, a visibility-preserving drone trajectory, or a full camera pose in $SE(3)$ .

2. Eva-Tracker as a data-driven event-camera feature tracker

In “Data-driven Feature Tracking for Event Cameras,” Eva-Tracker is the paper’s first data-driven feature tracker for event cameras (Messikommer et al., 2022). The pipeline is hybrid. A feature is first detected in a grayscale reference frame as a template patch $\mathbf{P}_0$ . For each later timestep $t_j$ , the tracker extracts an event patch $\mathbf{P}_j$ around the previous feature location from the asynchronous event stream $E_j$ , predicts the relative feature displacement $\mathbf{\Delta \hat{f}_j}$ , updates the feature location, crops a new event patch at the new predicted position, and repeats. The grayscale frame supplies the stable appearance template; low-latency events supply the motion cue.

The network has two major parts: a feature network that processes each track independently, and a frame attention module that shares information across tracks in the same image. The feature network uses separate patch encoders for the grayscale template and the current event patch, both based on Feature Pyramid Networks. Their outputs are compared through a correlation map $C_j$ , and the result is refined by a second encoder containing standard convolutions and a ConvLSTM block with temporal state $F_j$ . The frame attention module then performs self-attention across tracks, motivated by the observation that points on the same rigid object should move in correlated ways. The paper writes the module as $C_j$ 0 with a final linear layer on $S_j$ predicting the displacement.

Training is supervised on synthetic data with a truncated $L_1$ loss over relative displacement. The truncation excludes out-of-patch cases, which stabilizes supervision when a prediction leaves the current patch. The method also uses online affine augmentation, warping the current event patch by random rotation, translation, and scaling parameters $\mathbf{P}_0$ 0, then mapping the predicted displacement back through the inverse warp. The training strategy is explicitly synthetic-to-real: supervised pretraining is performed on Multiflow with 30,000 feature tracks and ADAM at learning rate $\mathbf{P}_0$ 1, and fine-tuning on real event data uses a pose supervision loss for 700 optimization steps at learning rate $\mathbf{P}_0$ 2.

The self-supervision stage uses camera poses and triangulation. For each predicted track $\mathbf{P}_0$ 3, the tracker triangulates a 3D point $\mathbf{P}_0$ 4 from predicted 2D positions and known camera poses via direct linear transform, then reprojects the result to obtain a supervision signal. This requires a static scene and calibrated camera poses, obtainable from COLMAP or motion capture.

Quantitatively, the method reaches average Feature Age / Expected FA of $\mathbf{P}_0$ 5 on EC in zero-shot transfer, and $\mathbf{P}_0$ 6 on EDS, improving to $\mathbf{P}_0$ 7 after pose-based fine-tuning. The paper summarizes this as relative feature age improvement of up to 120%, increasing to 130% after adaptation. Latency is also emphasized: about 19.7 patches in 17 ms on EDS and about 14.2 patches in 13 ms on EC on an Nvidia Quadro RTX 8000. Important implementation details include SBT-Max as the dense event representation, five temporal bins per polarity giving 10 channels, 31×31 input patches, Harris corners for initialization, and no template updates during evaluation.

3. Eva-Tracker as ESDF-update-free, visibility-aware aerial tracking

In “Eva-Tracker: ESDF-update-free, Visibility-aware Planning with Target Reacquisition for Robust Aerial Tracking,” Eva-Tracker is a trajectory-planning framework for aerial target tracking (Lin et al., 13 Feb 2026). Its purpose is not to track image features or camera pose directly, but to keep a drone continuously observing a moving target while avoiding collisions and recovering quickly when the target is temporarily lost. The central claim is that visibility-aware planning can be made cheaper by eliminating repeated ESDF updates and replacing them with a precomputed Field of View ESDF, or FoV-ESDF.

The system has three coupled stages. First, it predicts the target trajectory by fitting a smooth Bézier trajectory $\mathbf{P}_0$ 8 over a time window $\mathbf{P}_0$ 9 from observed target positions $t_j$ 0, minimizing the integral of squared second derivative under interpolation constraints. Second, it generates a visibility-aware initial path. For each planning step, it chooses a candidate observation point at preferred distance $t_j$ 1 along the ray from the predicted target toward the tracker,

$t_j$ 2

checks whether the segment from $t_j$ 3 to the predicted target is occluded, and, if occluded, searches along a circular arc of radius $t_j$ 4 around the target until it finds the nearest occlusion-free point. This procedure also underwrites target reacquisition: even when the target disappears, waypoints continue to be generated around the predicted target position at the desired observation distance.

Third, it performs FoV-ESDF-guided trajectory optimization. The camera field of view is defined as a quadrilateral pyramid

$t_j$ 5

with $t_j$ 6 chosen as the optimal observation distance $t_j$ 7. FoV-ESDF stores the distance from points inside the camera’s FoV to the nearest FoV boundary and sets the value to zero outside the FoV. The field is constructed offline via a distance transform; values and gradients are obtained by interpolation. Dynamic obstacles are handled by transforming obstacle points into the tracker frame at each step, but the field itself is not recomputed.

The optimized trajectory is parameterized as a MINCO trajectory and solved with L-BFGS. The objective combines visibility, energy, and feasibility penalties,

$t_j$ 8

where $t_j$ 9 contains differentiable occlusion, observation-distance, and yaw-alignment terms. The visibility cost is written as

$\mathbf{P}_j$ 0

In simulation, Eva-Tracker achieves the best average tracking distance, $\mathbf{P}_j$ 1 m, a yaw angle error of $\mathbf{P}_j$ 2 rad, the lowest occlusion rate at 4.36%, and the lowest failure rate at 6.75%. The initial path generator takes only $\mathbf{P}_j$ 3 ms on average; the optimization stage takes $\mathbf{P}_j$ 4 ms, for total planning time of $\mathbf{P}_j$ 5 ms. Real-world experiments use a quadrotor with a Livox Mid-360 LiDAR and an onboard NVIDIA Jetson Orin NX, Faster-LIO for localization, YOLOv11 for person detection, and MonoLoco for monocular 3D pose estimation. The paper is explicit that the reacquisition mechanism is strongest for short-term target loss and that FoV-ESDF assumes a fixed camera field of view.

GS-EVT is described in the provided material as a system also referred to as “Eva-Tracker” in the query, but its formal name is “GS-EVT: Cross-Modal Event Camera Tracking based on Gaussian Splatting” (Liu et al., 2024). It addresses event-based camera tracking by aligning integrated event images with differential renderings from a Gaussian Splatting map built from ordinary RGB frames. The framing is explicitly cross-modal: mapping uses frame-based cameras, tracking uses event streams, and the bridge is a 3D Gaussian Splatting scene representation.

The measurement model begins from event generation in log-intensity space,

$\mathbf{P}_j$ 6

Events are grouped into adaptive keyframes once a minimum event count is reached, producing an integrated intensity-change image $\mathbf{P}_j$ 7. The timestamp of each keyframe is set to the midpoint of the earliest and latest event timestamps in the group.

A key contribution is the pose parametrization using a reference pose plus first-order dynamics. Let $\mathbf{P}_j$ 8 be the camera state at the center of the event-accumulation interval. With linear velocity $\mathbf{P}_j$ 9, angular velocity $E_j$ 0, and accumulation duration $E_j$ 1, the boundary poses are

$E_j$ 2

The rendered event-like signal is then

$E_j$ 3

and tracking minimizes the discrepancy between normalized $E_j$ 4 and normalized $E_j$ 5. The optimization variables are pose correction $E_j$ 6 and velocity terms $E_j$ 7.

Optimization proceeds in a staggered coarse-to-fine scheme. In the coarse stage, event polarity is ignored and matching becomes polarity-free edge registration. In the fine stage, original polarities are restored and pose and velocity are optimized jointly. The implementation also uses Gaussian blur, an image pyramid, and event-region masking to stabilize gradients.

The method is evaluated on the VECtor dataset and on a self-collected dataset with an RGB-based map-building trajectory, a DAVIS346 event camera for tracking, and OptiTrack ground truth. The main metric is Absolute Trajectory Error in position and orientation. On VECtor, GS-EVT is reported to be 2–10× more accurate than EVPT and roughly on par with EVT on some sequences. Example results at 100% sequence length include about $E_j$ 8 cm / $E_j$ 9 on sofa_normal, about $\mathbf{\Delta \hat{f}_j}$ 0 cm / $\mathbf{\Delta \hat{f}_j}$ 1 on robot_normal, and about $\mathbf{\Delta \hat{f}_j}$ 2 cm / $\mathbf{\Delta \hat{f}_j}$ 3 on desk_normal. On self-collected data, the paper reports $\mathbf{\Delta \hat{f}_j}$ 4 cm / $\mathbf{\Delta \hat{f}_j}$ 5 on desk_seq4, $\mathbf{\Delta \hat{f}_j}$ 6 cm / $\mathbf{\Delta \hat{f}_j}$ 7 on keyboard_seq4, and $\mathbf{\Delta \hat{f}_j}$ 8 cm / $\mathbf{\Delta \hat{f}_j}$ 9 on helmet_seq1.

5. Comparative technical structure

A direct comparison shows that the three usages of Eva-Tracker solve different estimation problems under different observability assumptions (Messikommer et al., 2022, Lin et al., 13 Feb 2026, Liu et al., 2024). The event-camera feature tracker estimates local feature displacement conditioned on a grayscale template and asynchronous event patches. The aerial Eva-Tracker estimates a visibility-preserving drone trajectory conditioned on target motion prediction, LiDAR-derived obstacle geometry, and camera field-of-view constraints. GS-EVT estimates camera pose and first-order motion by photometrically aligning integrated event images to a frame-built Gaussian Splatting map.

The learning-optimization split is equally sharp. The feature-tracking Eva-Tracker is a learned model with FPN encoders, ConvLSTM temporal state, and a frame attention module that performs self-attention across tracks. The aerial Eva-Tracker is primarily an optimization framework: Bézier-based target prediction, visibility-aware initial path generation, FoV-ESDF-based differentiable objectives, MINCO parameterization, and L-BFGS trajectory optimization. GS-EVT is also optimization-based, but in a different sense: it is a differentiable geometric optimizer over pose and velocity, with analytically derived Jacobians propagated through Gaussian Splatting rendering.

Their robustness mechanisms are task-specific. In the event-camera feature tracker, robustness comes from synthetic-to-real training, shared context across feature tracks, and pose-based self-supervision through triangulation. In the aerial planner, robustness comes from FoV-specific visibility geometry, observation-distance control, and explicit target reacquisition without a separate recovery mode. In GS-EVT, robustness comes from realistic view-dependent rendering by 3D Gaussian Splatting and from a coarse-to-fine optimization schedule that moves from polarity-free alignment to polarity-aware refinement. This suggests that “Eva-Tracker” is better understood as a family of task-local design patterns than as a single reusable architecture.

6. Terminological boundaries and adjacent systems

The label should not be conflated with several nearby event-tracking systems that occupy adjacent but distinct problem spaces. EvHand-FPV is an efficient event-camera-based 3D hand tracking framework for egocentric first-person-view settings and is described as “not exactly the same by name,” even though it is a strong match if the intended query is an egocentric event-based hand tracker (Xu et al., 17 Sep 2025). FE-TAP is an image-event fusion point tracker that combines contextual information from image frames with the high temporal resolution of events for high-frame-rate point tracking (Liu et al., 2024). ETAP is the first event-only tracking-any-point method and extends TAP-style global context modeling to event cameras with a feature-alignment loss for motion-robust descriptors (Hamann et al., 2024). eMoE-Tracker, by contrast, is a one-stream transformer-based RGB-event tracker for object bounding-box tracking under motion blur, illumination variance, occlusion, and scale variation (Chen et al., 2024).

A common misconception is therefore to treat “Eva-Tracker” as a single benchmarked package with a fixed sensor stack and metric suite. The cited material points in the opposite direction. Depending on usage, the term can refer to feature age and expected feature age in event-feature tracking, tracking distance and occlusion rate in aerial visibility-aware planning, or absolute trajectory error in event-camera pose tracking. A plausible implication is that any technical discussion of Eva-Tracker should first specify the target variable—feature point, drone trajectory, or camera pose—before comparing architectures, datasets, or reported gains.