Match-Any-Events: A Cross-Domain Framework

• Match-Any-Events is a cross-domain framework enabling localization, comparison, and segmentation of arbitrary events from diverse data sources such as vision, sports, and CEP.
• It leverages techniques like temporal aggregation, transformer-based matching, and open-vocabulary segmentation to address challenges including wide-baseline variations and motion dependency.
• The concept supports applications from event-camera correspondence and sports indexing to complex event processing and fair allocation in repeated matching scenarios.

Match-Any-Events denotes a family of event-centric formulations in which a system must localize, compare, retrieve, segment, trigger on, or allocate with respect to arbitrary events rather than a narrow predefined target. In current arXiv literature, the phrase appears explicitly as the title of a zero-shot wide-baseline matcher for event cameras, and it also serves as a useful perspective for open-vocabulary event segmentation, full-match sports indexing, complex event processing, multi-event serverless triggering, and repeated matching under prefix-wise fairness constraints (Zhang et al., 20 Apr 2026).

1. Conceptual scope and event units

Across the literature, the object being “matched” varies substantially. In event-camera vision, the primitive object is an asynchronous event stream or an event-derived region. In sports analytics, it can be a match segment, a rally, a hit event, a shot caption, or a sequence of possession events. In complex event processing, it is a typed stream event participating in a composite pattern. In repeated allocation, it is a round-wise matching whose quality is evaluated over prefixes of time.

Area	Event unit	Match objective
Event cameras	Stream events or event regions	Correspondence or open-vocabulary segmentation
Sports analytics	Shots, rallies, possessions, match segments	Indexing, captioning, forecasting, valuation
CEP and serverless	Typed stream events	Pattern detection or trigger firing
Repeated matching	Round-wise allocations	Maximin fairness at final time or every prefix

This heterogeneity is not superficial. It changes the representation, supervision, and evaluation protocol. Event-camera work emphasizes correspondence fields, confidence maps, and wide-baseline robustness; sports work emphasizes temporal hierarchy, dense annotations, or predictive signatures; CEP work emphasizes denotational semantics and automata; repeated matching work emphasizes bottleneck utility over time. A plausible implication is that “Match-Any-Events” is best understood as a cross-domain design goal rather than a single model class.

2. Event-camera correspondence across wide baselines

The direct use of the term is "Match-Any-Events: Zero-Shot Motion-Robust Feature Matching Across Wide Baselines for Event Cameras," which studies semi-dense wide-baseline feature matching for event cameras under varying motion profiles in a zero-shot cross-dataset setting. Each event is represented as $e_i=(x_i,y_i,t_i,p_i)$, the input is converted into a logarithmically windowed event voxel, and the matcher combines a Temporal Aggregation Transformer, Sparsity-aware Event Token Selection, a ViT backbone followed by DPT, and a coarse-to-fine matching stage with dual-softmax, Mutual Nearest Neighbor selection, and local expected-coordinate refinement in a $3\times 3$ window. To support wide-baseline generalization, the work introduces E-MegaDepth with approximately 3 million pairs of event streams and ECM as a real synchronized event-image dataset. It reports a 37.7% improvement over the previous best event feature matching methods, ECM event-to-event performance of AUC@5° $54.61$ and precision $68.90$, EDS pose-estimation AUC $40.4/56.2/68.8$, runtime of 49 ms for a $350\times630$ event pair on an RTX 4080 GPU, and a 21.5% reduction in spatial-attention FLOPs from SETS (Zhang et al., 20 Apr 2026).

A central technical issue in this setting is motion dependence. Event appearance changes with motion, integration interval, and contrast thresholding, so the same scene point can generate substantially different event patterns across views. The paper addresses this by learning multi-timescale features and by separating spatial and temporal attention, reducing complexity from $\mathcal{O}((THW)^2)$ to $\mathcal{O}(T(HW)^2 + HW\,T^2)$. The reported motion study also shows that all methods degrade below 20 ms because of insufficient texture, while the proposed matcher maintains performance across a broad interval range and shows almost no degradation as interval increases (Zhang et al., 20 Apr 2026).

3. Open-vocabulary event segmentation and semantic grounding

A complementary event-camera interpretation of Match-Any-Events appears in "Segment Any Events with Language," which introduces SEAL as the first Semantic-aware Segment Any Events framework for Open-Vocabulary Event Instance Segmentation. The model takes event data, a visual prompt, and language, and supports both event segmentation and open-vocabulary mask classification at multiple levels of granularity, including instance-level and part-level. The main architecture combines a pretrained EventSAM backbone, a SAM mask decoder, Multimodal Hierarchical Semantic Guidance, and a Multimodal Fusion Network with a Backbone Feature Enhancer, Spatial Encoding, and Mask Feature Enhancer. In the supplementary formulation, mask generation is written as $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$, while semantic training is driven by cosine-alignment distillation to image-derived and text-derived targets across semantic, instance, and part levels (Lee et al., 30 Jan 2026).

The framework is notable because it transfers image-language supervision into the event domain without requiring dense event labels. Hierarchical visual guidance comes from image masks and CLIP image features, while hierarchical text guidance comes from MLLM-generated captions encoded by the CLIP text encoder. This yields a prompt-conditioned region representation that is semantically richer than class-agnostic EventSAM masks alone. Quantitatively, SEAL reports DDD17-Ins point AP $32.3$ and box AP $3\times 3$0, DSEC11-Ins point AP $3\times 3$1 and box AP $3\times 3$2, and DSEC-Part point AP $3\times 3$3 and box AP $3\times 3$4; total inference time is 22.28 ms and total parameter count is 99.1M. The appendix introduces SEAL++, a prompt-free variant with a class-agnostic detection branch, reporting DSEC-Detection generic EIS AP $3\times 3$5 at 24.12 ms (Lee et al., 30 Jan 2026).

This work makes the phrase “match any events” unusually literal: event regions can be segmented and then semantically aligned to arbitrary text at coarse, fine, or part-level granularity. The main paper remains prompt-conditioned, so it is not a text-only dense retrieval system. Still, a plausible implication is that promptable open-vocabulary segmentation is one of the clearest operational realizations of Match-Any-Events in event sensing.

4. Full-match sports understanding and arbitrary event access

In sports, the phrase functions less as a title than as a perspective on whole-match event access. "BFMD: A Full-Match Badminton Dense Dataset for Dense Shot Captioning" introduces the first Badminton Full Match Dense dataset, built from 19 full-length broadcast matches and explicitly preserving the complete match timeline. It covers 20.32 hours, 1,687 rallies, 16,751 hit events, 795 replay segments, 52 Hawk-Eye challenge segments, 419 net-hit events, and 1,556 shuttle-landings. The hierarchy has three temporal levels: match segments, rally events, and dense rally-level multimodal annotations. Each of the 16,751 hit events is associated with a corresponding caption, making the hit event or shot the atomic unit while retaining rally and match context. The benchmark model is a VideoMAE-based multimodal captioning framework with Semantic Feedback; on the singles subset it reports BLEU-4 $3\times 3$6, METEOR $3\times 3$7, ROUGE-L $3\times 3$8, and CIDEr $3\times 3$9, and the full multimodal model reaches BLEU-4 $54.61$0, METEOR $54.61$1, ROUGE-L $54.61$2, and CIDEr $54.61$3 (Ding et al., 26 Mar 2026).

The significance of BFMD for Match-Any-Events is architectural rather than benchmark-defined. The paper explicitly focuses on caption generation and does not perform event detection, event retrieval, or arbitrary-event search over full matches. Yet the dataset directly supports finding temporally localized shot events within full matches, dense temporal localization of rally boundaries and interruption segments, event sequence modeling across rallies and over the whole match, shot-level caption generation, and shot-semantic indexing. It indirectly enables event matching across matches, event retrieval given text or example events, cross-rally retrieval, and tactic retrieval, because all events remain in chronological order within complete broadcast videos and are frame-aligned (Ding et al., 26 Mar 2026).

Related sports papers broaden this match-any interpretation. "Visual analytics for team-based invasion sports with significant events and Markov reward process" values any kind of event with a continuous parameter space by extracting 202,168 significant events from 611 matches, representing a state as time, event type, location, side, and score, and solving a Markov reward process with fitted-value iteration; random forest yields RMSE $54.61$4 versus $54.61$5 for linear regression (Zhao et al., 2019). "Effect of Key Match Events on Football Passmaps" analyzes halftime, first goal, and first dismissal on 789 parsed matches and over 9000 unique networks, finding that first goal and first dismissal yield the clearest structural effects, whereas halftime shows no conclusive global event effect (Stropnik et al., 2021). "Detecting key Soccer match events to create highlights using Computer Vision" treats foul, corner kick, goal, and penalty kick as frame-level detection targets; Faster R-CNN with ResNet50 reports class accuracy $54.61$6 versus $54.61$7 with VGG16, and a 23-minute video is reduced to 4:50 minutes of highlights (Darapaneni et al., 2022).

5. Predictive signatures and few-shot event matching

A second major line of work represents events through predictive or metric structure rather than explicit localization. "Transformer-Based Neural Marked Spatio Temporal Point Process Model for Football Match Events Analysis" models football events as tuples $54.61$8, where $54.61$9 is interevent time, $68.90$0 is one of 20 pitch zones, and $68.90$1 is one of 5 action classes: pass, dribble, cross, shot, or possession end. The model factorizes next-event prediction as $68.90$2, uses a Transformer history encoder over the last 40 events, and finds the best dependency order to be $68.90$3. It reports validation total loss $68.90$4, tied-best interevent RMSE $68.90$5, and introduces HPUS, whose average correlates with final ranking at $68.90$6 and with goals and xG at $68.90$7 (Yeung et al., 2023).

"Forecasting Events in Soccer Matches Through Language" reframes next-event prediction as next-token generation over a structured event language. The raw schema has 11 fields, but the generated chain is effectively Event Type, isGoal, isAccurate, isHomeTeam, TimeElapsed, X, and Y, produced one token at a time by an MLP over a shared vocabulary. With $68.90$8 previous events as context, the model reports event-type ACC $68.90$9 and F1 $40.4/56.2/68.8$0, accurate ACC $40.4/56.2/68.8$1 and F1 $40.4/56.2/68.8$2, TimeElapsed MAE $40.4/56.2/68.8$3, X MAE $40.4/56.2/68.8$4 with $40.4/56.2/68.8$5, and Y MAE $40.4/56.2/68.8$6. The same generative backbone is then used for situational expected-goals maps, momentum curves, long-term match probabilities, and VAEP-like action valuation (Mendes-Neves et al., 2024).

Few-shot work approaches Match-Any-Events as generalized similarity learning. "Learning to match transient sound events using attentional similarity for few-shot sound recognition" replaces clip-level pooled comparison with an attention-guided segment similarity

$40.4/56.2/68.8$7

or equivalently $40.4/56.2/68.8$8, thereby emphasizing short event-bearing regions. The reported relative improvement in 5-shot 5-way accuracy ranges from $40.4/56.2/68.8$9 to $350\times630$0 on ESC-50 and from $350\times630$1 to $350\times630$2 on noiseESC-50 (Chou et al., 2018). "Extensively Matching for Few-shot Learning Event Detection" brings the same idea to NLP event detection by adding support-to-support structure through

$350\times630$3

and

$350\times630$4

used together with the standard episodic query loss in an $350\times630$5-way $350\times630$6-shot setting with a NULL class. On ACE-2005, one of the strongest gains is LSTM Proto $350\times630$7 in 10+1-way 10-shot, and LSTM Proto+Att $350\times630$8 under the same setting (Lai et al., 2020).

Taken together, these papers treat “matching” as either a predictive signature over event histories or a learned metric over support and query events. The unifying theme is that arbitrary-event comparison becomes easier when the representation preserves temporal context, local salience, or support-set geometry.

6. Declarative event matching, automata, and trigger semantics

Complex event processing provides the most formal account of Match-Any-Events. "Symbolic Automata with Memory: a Computational Model for Complex Event Processing" introduces Register Match Automata, defined as $350\times630$9, with finite states, registers, and transitions of the form $\mathcal{O}((THW)^2)$0. Unlike symbolic automata, RMA allow formulas to be applied not only to the last element read from the input string, but to multiple elements stored in registers; unlike register automata, they allow arbitrary formulas besides equality predicates. A run yields a match by collecting the stream positions marked $\mathcal{O}((THW)^2)$1, and the paper gives a systematic construction from bounded core-CEPL expressions with $\mathcal{O}((THW)^2)$2-ary constraints. RMA are closed under concatenation, union, Kleene$\mathcal{O}((THW)^2)$3, and determinization, but not under complement; they are not determinizable as recognizers in general, although bounded windowing restores determinization in that view (Alevizos et al., 2018).

A more operational trigger-language view appears in "Multi-Event Triggers for Serverless Computing." The trigger grammar is $32.3$2 so a rule is either a leaf $\mathcal{O}((THW)^2)$4 or a binary composition $\mathcal{O}((THW)^2)$5 or $\mathcal{O}((THW)^2)$6. This supports count-based triggers, joins over heterogeneous event types, and match-any behavior through OR. Example rules include 3:a, AND(2:a,2:b), OR( AND(6:temperature,6:wind), AND(1:temperature,1:motion) ), and OR( AND(5:packetLoss,1:temperature), 1:powerConsumption ). Internally, each trigger maintains one trigger set for each event type; when a rule is fulfilled, the satisfying events are removed and sent to the function. In the incident-detection use case, this reduces median event–invocation latency by 62.5%, and the prototype reaches 313,154.81 requests/s on four nodes (Carl et al., 27 May 2025).

The commonality between RMA and METs is that both treat “matching” as satisfaction of structured predicates over streams, not merely nearest-neighbor similarity. RMA emphasize denotational semantics and finite-memory automata; METs emphasize platform-level trigger execution with count, AND, and OR operators. In both cases, Match-Any-Events means matching composite patterns rather than isolated events.

7. Fairness, recommendation, and recurring limits

In repeated allocation, Match-Any-Events appears as prefix-wise fairness over repeated matchings. "Fairness in Repeated Matching: A Maximin Perspective" studies sequences of injective round-wise matchings $\mathcal{O}((THW)^2)$7 over $\mathcal{O}((THW)^2)$8 rounds, with cumulative utility

$\mathcal{O}((THW)^2)$9

and bottleneck value

$\mathcal{O}(T(HW)^2 + HW\,T^2)$0

A sequence is optimal at round $\mathcal{O}(T(HW)^2 + HW\,T^2)$1 if $\mathcal{O}(T(HW)^2 + HW\,T^2)$2, and anytime optimal up to round $\mathcal{O}(T(HW)^2 + HW\,T^2)$3 if $\mathcal{O}(T(HW)^2 + HW\,T^2)$4 for all $\mathcal{O}(T(HW)^2 + HW\,T^2)$5. ERM is NP-complete and APX-hard even when $\mathcal{O}(T(HW)^2 + HW\,T^2)$6 and utilities are ternary in $\mathcal{O}(T(HW)^2 + HW\,T^2)$7. Exact anytime optimality always exists for $\mathcal{O}(T(HW)^2 + HW\,T^2)$8, may fail for $\mathcal{O}(T(HW)^2 + HW\,T^2)$9, and deciding its existence is coNP-hard. The paper therefore gives additive approximations, including

$$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$0

for the anytime case (Lim et al., 6 Oct 2025).

A text-centric interpretation appears in "A novel recommendation system to match college events and groups to students." There, items are described in free text, user models are built from a few keywords expanded into weighted near-synonyms, item vectors are computed over the user vocabulary with field-weighted modified tf-idf, and cosine similarity is used for ranking. The modified tf-idf is binary: $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$1 and the similarity is cosine. On the reported evaluation, stemming improves precision to $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$2 and accuracy to $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$3, removing final $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$4 after stemming yields precision $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$5 and accuracy $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$6, binary tf-idf reaches precision $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$7 and accuracy $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$8, and user-model updating increases these to $$M = M_{\mathrm{sam}}(F_{\mathrm{evt}}(I_{\mathrm{evt}}), P_{\mathrm{sam}}(P))$$9 and $32.3$0 (Qazanfari et al., 2017).

Several recurring limits cut across the literature. BFMD provides dense full-match annotations but explicitly focuses on caption generation and does not benchmark retrieval or event detection (Ding et al., 26 Mar 2026). Match-Any-Events for event cameras depends on large-scale synthetic supervision and still degrades below 20 ms because of insufficient texture (Zhang et al., 20 Apr 2026). SEAL’s main formulation requires visual prompts, and the paper notes difficulties in low-event-rate regions and domain gaps to indoor scenes (Lee et al., 30 Jan 2026). RMA need bounded expressions and, for recognizer-style determinism, bounded windows (Alevizos et al., 2018). Exact anytime fairness in repeated matching may not exist once $32.3$1 (Lim et al., 6 Oct 2025). These limitations do not negate the Match-Any-Events agenda, but they define its present boundary: arbitrary-event access is increasingly feasible when event structure is hierarchically annotated, semantically aligned, finitely representable, or prefix-regularized, yet it remains constrained by memory, supervision, and evaluation design.