Visual Drift Detection (VDD)

• Visual Drift Detection (VDD) is a methodology that identifies, categorizes, and quantifies temporal changes in process logs and high-dimensional visual data using statistical analysis and visual representations.
• It integrates techniques from process mining and computer vision—such as clustering, change point detection, and GAN-based analysis—to capture both abrupt and gradual drifts with high accuracy.
• The approach supports actionable insights through tailored visual analytics and quantitative drift measures, enabling effective model retraining, process adaptation, and enhanced MLOps.

Visual Drift Detection (VDD) denotes the identification, characterization, and interpretation of change over time in observed data or inferred behavior, with the term used in two closely related but distinct ways. In process mining, VDD is a named technique for detecting, categorizing, drilling down into, and quantifying process drifts from event logs by combining declarative process constraints, clustering, change point detection, and visual analytics. In computer vision and visual data analytics, the same expression is used more broadly for methods that detect drift in high-dimensional image or video data, image sequences, and deployed perception systems, often under changing weather, lighting, viewpoint, season, or other contextual factors (Yeshchenko et al., 2019, Suprem et al., 2020, Basak et al., 17 Jun 2025, Cobb et al., 2022).

1. Scope and meanings of VDD

The process-mining formulation treats drift as a change in business processes as observed in event logs. Its central requirements are drift identification, categorization, drill-down, and quantification, and it operationalizes these requirements through mined declarative constraints and dedicated visualizations (Yeshchenko et al., 2019).

The vision-oriented formulation treats drift as a change in the data generating process or in the distribution of visual observations. In this setting, drift may arise from snow, weather, lighting, point of view, synthetic-to-real transfer, season, background shift, occlusion, or sensor degradation, and it affects the accuracy, calibration, and reliability of deployed models (Suprem et al., 2020, Dadboud et al., 2024, Li et al., 13 Dec 2025).

A further statistical formulation emphasizes that recent deployment batches are often not i.i.d. samples from historical reference data because time-induced correlation and other contextual factors are permitted to change. Under that view, the objective is not always unconditional two-sample testing, but sometimes testing for differences in distributions conditional on context (Cobb et al., 2022).

Setting	Drift object	Representative mechanisms
Process mining	Executed business-process behavior in event logs	Declare constraints, clustering, PELT, DriftMap, DriftChart (Yeshchenko et al., 2019)
Video analytics	High-dimensional image distributions	DA-GAN latent space, density bands, KL divergence, specialization and selection (Suprem et al., 2020)
Image-sequence monitoring	Drift patterns in jump location curves	Oblique-axis regression tree, prediction error, CUSUM chart (Basak et al., 17 Jun 2025)
Vision benchmarking	Longitudinal or domain-shifted imagery	CNN-VAE drift metrics, background-wise metrics, MCDO-map (Li et al., 13 Dec 2025, Dadboud et al., 2024)

2. VDD in process mining

The named VDD technique begins by chronologically sorting an event log $L$ and partitioning it into windows of fixed size $w$, shifted by step size $s$, with the number of windows given by

$$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$

Within each window, Declare constraints are mined over the process activity alphabet $\Sigma$, and the confidence of each constraint is tracked over time, producing a multivariate time series

$D = \{T_1, T_2, \dots, T_n\}.$

This transforms process drift analysis into time-series analysis on constraint confidence trajectories (Yeshchenko et al., 2019).

To support drill-down and categorization, VDD clusters constraints whose confidence values evolve similarly over time. One formulation uses hierarchical clustering with correlation as the distance metric and the “weighted” linkage method; a later system description also reports hierarchical clustering with Ward linkage and Euclidean/correlation distance, together with constraint reduction through subsumption handling. Change point detection is then applied globally and within clusters using the Pruned Exact Linear Time (PELT) algorithm with a kernel-based cost function, allowing overall drifts and cluster-specific drifts to be separated (Yeshchenko et al., 2019, Yeshchenko et al., 2020).

The visual layer is integral rather than ancillary. DriftMap plots constraints against time, colors points by confidence, groups constraints by cluster, and overlays vertical lines at change points. DriftChart plots average cluster confidence over time with detected change points, supporting qualitative recognition of sudden, gradual, incremental, reoccurring, or seasonal behavior. The extended Directly-Follows Graph (eDFG) augments workflow structure with declarative constraints, helping localize drift to specific process segments (Yeshchenko et al., 2020).

VDD also defines quantitative measures for prioritization. For a cluster $C$, the erraticity measure is built from

$$\Delta(T_i) = \sum_{j=1}^{N_{win}-1} \delta(T_{i,j}, T_{i,j+1}), \qquad \mathcal{E}(C) = \sum_{i=1}^{|C|} \sqrt{1 + (\Delta(T_i))^2},$$

so that clusters with higher $\mathcal{E}$ are treated as more erratic. On synthetic logs, the technique achieved an average F-score of $0.96$; one evaluation reported default F-score $w$0 for most logs and $w$1 for the Loop log, with cluster-based detection raising the Loop F-score to $w$2. On real-world logs, VDD detected all drifts found by ProDrift and additionally revealed more fine-grained and cluster-specific drifts plus outliers. A user study with 12 process mining experts reported usefulness and ease-of-use scores between 5 and 6 on a 7-point scale (Yeshchenko et al., 2019, Yeshchenko et al., 2020).

3. Representation-based VDD for visual data

In video analytics, ODIN performs automated drift detection by learning the distribution of high-dimensional images with a Dual-Adversarial GAN (DA-GAN), composed of an encoder $w$3, decoder $w$4, latent discriminator $w$5, and image discriminator $w$6. New samples are projected into latent space, assigned to existing clusters when they fall within a cluster’s density band, and otherwise accumulated in a temporary cluster. A drift is signaled when the temporary cluster stabilizes, measured through the Kullback-Leibler divergence

$w$7

with promotion to a permanent cluster when $w$8. ODIN then invokes recovery through YOLO-SPECIALIZED or YOLO-LITE models and model selection through policies such as KNN-W and A-BM. On Berkeley DeepDrive video, the combined system reported $w$9 detection accuracy, $s$0 higher throughput, and $s$1 smaller memory than a static YOLO baseline (Suprem et al., 2020).

A different line of work frames drift detection as conditional rather than unconditional distribution testing. Context-Aware Drift Detection introduces a context variable $s$2 and tests whether $s$3 for deployment contexts, using maximum conditional mean discrepancies. Its conditional MMD is

$s$4

and the deployment-conditional statistic is the Average Distributional Treatment Effect on the Treated (ADiTT). This formulation is explicitly designed to be insensitive to prevalence shifts in permitted context while remaining sensitive to within-context drift, and the empirical study reports applicability to ImageNet-scale vision problems (Cobb et al., 2022).

For drone detection under domain shifts, the DrIFT dataset isolates point-of-view shift, synthetic-to-real shift, season shift, adverse weather shift, and background shift. It contributes background segmentation maps and the uncertainty metric MCDO-map, which has lower postprocessing complexity than traditional methods and is used in an uncertainty-aware unsupervised domain adaptation method. The reported empirical behavior includes Pearson correlation greater than $s$5 between MCDO-map and KL-divergence between source and target feature map distributions, together with superior performance to state-of-the-art unsupervised domain adaptation techniques (Dadboud et al., 2024).

4. Visual analytics and temporal characterization

DriftVis addresses concept drift through coordinated visual analytics rather than detector output alone. It combines a distribution-based detector with a streaming scatterplot, a drift degree line chart, density-difference views, and coordinated prediction-level displays. Drift quantification is based on the energy distance

$s$6

where $s$7, $s$8, and $s$9 aggregate within- and cross-sample Euclidean distances. Drift is measured per cluster under an incremental Gaussian Mixture Model and then aggregated into a global drift degree. The system is organized around three phases—detection, examination, and correction—and supports analyst actions such as selecting samples, merging clusters, and building new base learners. On synthetic datasets with 99 simulated drifts each, DriftVis detected 97.9/99 drifts on one benchmark and was competitive on variance-only drifts; case studies on weather prediction and text classification showed visually guided adaptation improving model accuracy (Yang et al., 2020).

Sequential image monitoring emphasizes gradual drift patterns rather than only abrupt changes. The oblique-axis regression tree method models grayscale images $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$0 and is explicitly designed to preserve and track jump location curves (JLCs). It fits an oblique tree to a moving window of past images, predicts the current image through leaf-node averaging, computes an aggregate prediction error $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$1, and updates a CUSUM statistic

$$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$2

Under the null hypothesis of no change in JLC drift pattern, the charting statistic is asymptotically standard normal; under a change in at least one JLC’s drift pattern, its mean and variance increase. The method is reported to distinguish gradual drift from abrupt change by cumulative small error increases versus immediate large error jumps, and simulation results show low detection delay, often 1–4 for moderate drift changes, while retaining in-control Average Run Length near the nominal target (Basak et al., 17 Jun 2025).

5. Datasets, benchmarks, and drift metrics

The Bristol streetlight dataset provides a longitudinal benchmark for visual monitoring and spatio-temporal drift detection. It contains over 526,000 images captured hourly by 22 fixed-angle cameras from 2021 to 2025, together with timestamps, GPS coordinates, and device identifiers. Its reference implementation trains one CNN-VAE per node and lighting condition, for 44 models total, and defines two per-sample drift metrics: relative centroid drift, which measures latent-space deviation from a baseline quarter, and relative reconstruction error, which measures normalized image-domain degradation. Experimental findings distinguish “Type-0” and “Type-1” visibility scenarios, report higher drift variance for larger latent spaces such as $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$3, and recommend moderate latent size $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$4 or $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$5 as balancing expressiveness with drift robustness (Li et al., 13 Dec 2025).

The DrIFT benchmark complements this longitudinal setting with controlled cross-domain variation. It contains 14 distinct domains, over 47,000 annotated image frames, balanced training and validation splits for almost all domains, bounding boxes, and background segmentation maps labeling pixels as sky, tree, or ground. This enables background-wise metrics such as $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$6, background-aware uncertainty analysis, and evaluation of “unseen background” effects under domain shift (Dadboud et al., 2024).

These resources embody two different benchmark philosophies. The streetlight corpus emphasizes naturally evolving long-term drift with rich metadata and per-sample unsupervised drift scores. DrIFT emphasizes controlled domain-shift axes and explicit background annotations for evaluation of uncertainty-aware adaptation. Together they indicate that VDD benchmarking has expanded from one-off drift detection toward long-horizon monitoring, context splitting, and drift-aware MLOps (Li et al., 13 Dec 2025, Dadboud et al., 2024).

6. Boundaries, adjacent problems, and common confusions

A recurrent source of ambiguity is that “visual drift” does not always mean dataset or process distribution shift. In industrial anomaly detection, unsupervised Visual Defect Detection is a distinct problem concerned with defect detection and localization from scarce defect samples. Score-DD trains score-based generative models on normal data, argues that reconstruction loss is unreliable because normal pixels also change during reconstruction, and instead uses the time-dependent gradient value of the normal data distribution as a metric, together with a $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$7 scales approach that reduces inference cost from full denoising to $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$8. This suggests a conceptual boundary between defect localization and temporal drift monitoring, even though both areas share anomaly-oriented and unsupervised machinery (Teng et al., 2022).

In visual odometry and visual-inertial odometry, drift denotes cumulative pose error. A feedforward Drift Reducing Neural Network for monocular visual odometry uses an 11-dimensional input composed of an initial orientation increment and statistical moments of feature motion, and reports orientation RMSE reductions of 66–76% and translation RMSE reductions of 65–85% on KITTI. PGD-VIO instead uses point and plane features in an EKF and adds a graph-based drift detection strategy that searches overlapping and identical structures in the plane map, then suppresses cumulative drift through equality constraints on matched planes, without loop closure or global bundle adjustment. In offloaded VIO for VR/XR, adversarial slow-pose drift is treated as a temporal anomaly: an autoencoder trained on attack-free sessions classifies incoming slow poses by reconstruction MSE, drops anomalous poses, and runs in under 5 ms per slow-pose window; at 25–50% spoofing, the defense reduces trajectory errors by over $$N_{win} = \left\lfloor \frac{|L| - w}{s} \right\rfloor .$$9, while at 75% spoofing it still suppresses local jitter although global drift accumulates with forced passes (Wagih et al., 2022, Zhang et al., 2024, Saha et al., 8 Sep 2025).

A second misconception is that any distributional change should trigger an alarm. Context-aware testing argues explicitly against that view by making detectors insensitive to permitted changes in context prevalence, while the streetlight study recommends adaptive sensitivity because day/night and visibility regime materially alter drift behavior. ODIN, DrIFT, and the streetlight benchmark all further imply that detection alone is insufficient when deployment demands retraining, specialization, or drift-triggered model adaptation (Cobb et al., 2022, Li et al., 13 Dec 2025, Suprem et al., 2020, Dadboud et al., 2024).