Perceive Once, Decide Per Query: CogVis Splits Change Detection by Decision Scope

TL;DR for operators

When the same before-and-after imagery must answer several semantic questions, rerunning the full visual analysis for every query wastes computation that does not actually depend on the query.

CogVis separates those decisions by scope. It computes category-independent temporal change evidence once, then repeats only the semantic calibration and candidate verification that must depend on the requested category. The paper reports 28.50% higher inference throughput than the next-fastest compared method; with 10 queries, CogVis takes 10.03 seconds versus 12.56–956.60 seconds for the category-wise baselines evaluated.

The accuracy evidence is also broad: CogVis reports the strongest result on all seven target benchmarks, with improvements over the best prior reported result ranging from 0.05 to 8.17 percentage points. Controlled ablations suggest that the three stages are complementary rather than interchangeable.

For operators building multi-query monitoring systems, the more general design principle is to identify which expensive evidence can be shared across decisions and which thresholds or verification steps genuinely need to remain request-specific. The boundary is equally important: CogVis does not make the whole pipeline query-independent, does not establish robustness to arbitrary prompt wording, and does not validate real-world wildfire damage estimates.

The same scene should not require the same perception work ten times

A monitoring team may inspect one pair of satellite images for several questions: whether buildings changed, whether vegetation changed, or whether another requested category appeared or disappeared. The semantic questions differ, but the underlying before-versus-after imagery does not.

A straightforward pipeline can nevertheless repeat substantial visual processing for every question. That becomes increasingly expensive as the number of queried categories grows.

CogVis, proposed by Zijie Wang, Chen Zhong, and Wei He,¹ starts from a narrower question: which parts of change detection depend on the scene, and which depend on the query?

That distinction matters because the task is more demanding than identifying pixels that look different. In open-vocabulary change detection, a text-specified category determines what kind of change the system is looking for. Some temporal evidence—whether an area changed at all—can therefore be useful across many categories. But deciding whether that evidence corresponds to a particular requested category still requires query-specific reasoning.

CogVis reports 0.230 FPS and 4.82 GiB peak GPU memory in its RTX 3090 profiling, with throughput 28.50% higher than the next-fastest comparison. At 10 queries, its reported inference time is 10.03 seconds, compared with 12.56–956.60 seconds for category-wise baselines.

The speed advantage is not attributed to having the smallest model. It comes from avoiding repeated temporal perception.

CogVis assigns three decisions to three different scopes

The architecture implements this separation through three stages.

SCP uses frozen SAM3 features together with temporal differences, feature products, a low-rank Change Adapter, and a shallow change head. The result is a scene-level change prior: evidence about where meaningful temporal change is likely, without yet deciding which semantic category caused it.

This prior is reusable across queries. In the paper’s analysis, SCP raises ROC AUC from 0.79 for raw RGB differences to 0.96, supporting its role as more than a computational shortcut: it also produces a cleaner temporal signal.

SMC then asks the query-specific question. It combines semantic transition evidence, support from the shared temporal prior, stable-region suppression, and the change in category presence between the two timestamps. Conceptually, a strong response requires several pieces of evidence to agree rather than treating semantic similarity alone as sufficient.

But generating a score is only half the decision. The same raw response level may not imply the same operating point across different scenes and categories.

Reuse stops where calibration begins

This is where an overly broad reading of CogVis becomes misleading. Only temporal perception is shared. The semantic decision boundary is still adapted separately for each image-query pair.

SMC retrieves the 16 nearest response-threshold memories constructed from CA-CDD and uses them to obtain a threshold anchor. A learned Score Adapter then corrects that operating point for the current image and query, subject to a label-free lower bound and fixed threshold limits.

The supplementary calibration experiment is best interpreted as a mechanism test rather than a second headline benchmark. On SECOND and CLCD, a fixed global threshold averages 19.25 across the reported metrics. A memory-derived threshold anchor reaches 23.38. Adding the Score Adapter raises the mean to 25.19—5.94 points above the fixed threshold and 1.81 above memory retrieval alone.

That pattern supports two separate claims. Historical operating points contain useful information, but retrieving one is not enough. The decision boundary still benefits from correction conditioned on the current case.

The memory experiment also tests whether this benefit requires a large store. Using only 10% of the memory bank retains 94.9%, 97.7%, and 95.6% of the full-memory gains on SECOND, CLCD, and xBD respectively, while reducing storage from 525.2 MiB to 52.6 MiB. That is a sensitivity and efficiency result, not evidence that memory size is generally unimportant; it shows that substantial redundancy exists under this paper’s construction and benchmarks.

Region filtering works best after the upstream evidence is reliable

After thresholding, CogVis obtains connected candidate regions. ARF evaluates each candidate using a learned 19-dimensional description incorporating semantic, temporal, geometric, and calibrated-margin evidence.

Its behavior is deliberately one-sided: ARF can keep or remove an existing region. It cannot restore a true change region already lost upstream.

The component ablation makes the dependency among stages visible. Relative to the Vanilla system, the full SCP+SMC+ARF configuration improves SECOND from 19.36 to 32.48 mIoU, CLCD from 12.10 to 33.67 IoU, and xBD from 28.43 to 32.68 mIoU.

More revealing is what happens before the full system is assembled. Adding ARF after SCP+SMC improves those three benchmarks by 4.44, 4.76, and 1.82 points respectively. By contrast, ARF without reliable upstream temporal and semantic evidence is less consistent.

The region-cue ablation reinforces that interpretation. Removing temporal evidence causes the largest mean deterioration in ARF’s region-ranking metrics, 0.0519, compared with 0.0206 for geometry and 0.0133 for semantic evidence. ARF is therefore better understood as a verifier whose effectiveness depends on the evidence supplied to it, not as an independent cleanup module that can compensate for arbitrary upstream errors.

The benchmark results connect the architecture to a deployment trade-off

Across SECOND, SCSCD, CLCD, DSIFN, LEVIR-CD, WHU-CD, and xBD, CogVis reports the strongest metric among the compared methods, with margins over the strongest prior reported results ranging from 0.05 to 8.17 percentage points.

Its reported results are 32.48 and 25.36 mIoU on SECOND and SCSCD; 33.67, 48.50, 72.55, and 75.86 IoU on CLCD, DSIFN, LEVIR-CD, and WHU-CD; and 32.68 mIoU on xBD.

For Cognaptus, the broader inference is not that every multi-query AI system should copy this architecture. The transferable design question is narrower: what evidence is invariant across requests, and what decision must still be recalibrated for each request?

In an Earth-monitoring workflow, repeatedly recomputing the same temporal interpretation raises latency with little semantic benefit. Sharing it can reduce that repeated cost. But sharing the later decision boundary would remove exactly the category dependence the task requires.

This suggests a useful systems principle for multi-query perception: amortize reusable evidence production, not the decisions whose reliability depends on user intent.

The current evidence stops short of unrestricted operational deployment

Several boundaries materially constrain that interpretation.

First, SMC and ARF still execute per query. CogVis reduces repeated computation rather than eliminating query-dependent scaling.

Second, every benchmark uses a fixed prompt vocabulary across its full test split. The experiments therefore support cross-benchmark generalization under predetermined text categories, not robustness to arbitrary image-specific wording or prompt engineering.

Third, ARF cannot recover missed candidates. If upstream thresholding removes a genuine change region, downstream verification has nothing to reconsider.

Fine-grained semantic distinctions also remain difficult, including Playground and Structure categories and subtle damage grades such as minor damage.

Finally, the supplementary Palisades and Eaton wildfire application demonstrates qualitative spatial transfer across 2,501 tiles using the fixed model and xBD prompts without site-specific fine-tuning. But there are no local reference labels. The case therefore cannot establish local accuracy or support validated estimates of damage extent.

Separate reusable perception from request-specific judgment

CogVis’s strongest contribution is the way it assigns computation according to decision scope.

Scene-level temporal evidence is calculated once because the before-and-after relationship does not change when the user asks a different semantic question. The response threshold remains query-specific because score meaning does change with the image and requested category. Region verification stays candidate-specific because reliability depends on local evidence.

The seven-benchmark results and controlled ablations give that decomposition more support than an efficiency result alone would provide. They indicate that reuse, calibration, and verification address different failure modes and work most reliably in sequence.

For multi-query monitoring systems, this shifts architecture design away from asking how much of a pipeline can be made generic. The more precise decision is which computations remain valid when the query changes—and where reuse would begin to erase information the system still needs.

Cognaptus: Automate the Present, Incubate the Future.