acceptodds
Under review as a conference paper at ICLR 2027

WEAK QUERY CALIBRATION FOR MULTI-VIEW UNIQUE COUNTING UNDER UNRELIABLE OVERLAP PROXIES

Abstract

We introduce DAWSON, a weak-query calibration framework for multi-view unique counting with a small budget of aggregate shared-person answers. It treats each answer as sparse count-level supervision: per-view counts describe observations, while a queried pair supplies a shared-person cardinality without requiring cross-view identities. People visible in several cameras create duplicate mass, and appearance proxies weaken under occlusion, viewpoint changes, crop loss, and detector noise. Thus, the pair with the largest estimated overlap need not be the query that most improves the final union estimate. DAWSON combines per-view counts and pairwise overlap proxies into a fused unique-count estimate. Its expected union-error reduction (EER) rule evaluates candidate answers through an observable response surrogate and selects the query with the largest expected decrease in a calibration residual. The residual couples count consistency with uncertainty in unrevealed overlaps, so the weak-label budget follows the counting state. Each answer updates the fused estimate and ranking of remaining candidates. Selection uses observable counts, proxy similarities, and revealed responses, while keeping cross-view identities and the ground-truth unique count outside acquisition. Experiments on synthetic scenes and the static-camera WILDTRACK image-proxy bridge use annotation-simulated responses. With matched inputs, estimators, thresholds, and two-query budgets, EER yields lower paired MAE than Largest-overlap in all twelve degraded settings across four frozen encoders, with every paired 95% confidence interval below zero. The gap reaches −14.68 MAE for DINOv3 at Strong degradation. For YOLO26m, relative MAE reductions are 12.8%, 11.1%, and 13.2% at Mild, Medium, and Strong, and persist with noisy answers. Reliable-proxy controls retain Largest-overlap as a strong baseline, with same-input Max providing a zero-query reference for assessing weak-query calibration. These results establish residual-driven query selection as a proxy-regime-dependent mechanism for correcting overlap error under constrained budgets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi-view object counting, unique object counting, set-union estimation, active querying, weak supervision, overlap estimation, occlusion robustness

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_b9jLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DAWSON, a framework for multi-view unique counting that uses a limited number of shared-count queries to correct unreliable overlap estimates. It combines per-view counts with pairwise overlap proxies to estimate the total number of distinct people. The EER strategy uses a surrogate for the possible query responses and selects the camera pair expected to reduce the calibration residual the most.The paper evaluates this approach under both clean and degraded proxy conditions.

Critical strengths

The unique-count estimator in Eq. (2) is interesting. It provides a simple and intuitive way to estimate the number of unique individuals across multiple views.

Critical weaknesses
  1. The experiments allow human queries in every evaluation window. Would this become too costly in continuous use? In the Mild setting in Table 2, the best querying policy improves about 4.4 people. It is unclear whether this gain is worth the human effort, especially since answering a shared-count query may require checking identities across views.

  2. Similarly, The paper should also report end-to-end runtime, hardware, and the extra cost of EER query selection. It would be useful to see how these costs grow with the number of cameras and people.

Questions for authors

Refer to weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer b9jL

1.Comment: The experiments allow human queries in every evaluation window. Would this become too costly in continuous use? In the Mild setting in Table 2, the best querying policy improves about 4.4 people. It is unclear whether this gain is worth the human effort, especially since answering a shared-count query may require checking identities across views. 1.Reply: The submitted manuscript already reported the Mild ablation with two annotation-simulated answers per window. We add the matched zero-query Max reference and an offline human-answer pilot to make the effort--accuracy trade-off explicit. EER reduces MAE from 17.18 for No-query to 12.76, a 4.41-person difference; against the matched two-query Largest policy at 13.78, the improvement is 1.01 MAE. Zero-query Max reaches 0.67.

We evaluated the human effort in a separate strict-box OT pilot. One author answered all 42 pair-query tasks offline across two development windows before policy replay. Of 41 certain answers, 5 were exact and the answer MAE was 2.488. In windows 168 and 180, the selected answers required 55.574 and 39.149 seconds of recorded active work, respectively. EER and Largest selected the same pair set in each window, with reverse order in window 180; each had absolute count errors of 1.639 and 3.828 in the respective windows, while Max had error 1 in both.

These observations quantify recorded offline effort and answer accuracy,therefore they support a direct workload comparison without establishing that continuous querying is economically worthwhile in deployment.

2.Comment: Similarly, The paper should also report end-to-end runtime, hardware, and the extra cost of EER query selection. It would be useful to see how these costs grow with the number of cameras and people. 2.Reply: The submitted manuscript did not provide this runtime breakdown; we add a strict-box F0/OT measurement on an RTX 4090 D over 20 held-out windows (200--390) and three system repeats. Each four-policy group shared one model load; each policy used independent warm-up windows. Direct timing runs from the first image read to the strategy output, includes the simulated evaluator, and excludes model loading, warm-up, and final persistence. EER wall-clock medians for 2/3/5/7 views are 0.325/0.503/0.780/1.139 s. At matched inputs, EER minus Largest policy-time medians are 0.163/0.760/2.572/7.126 ms across these view counts.

The selector increment grows with candidate pairs (1, 3, 10, and 21 pairs for 2, 3, 5, and 7 views), while the full wall-clock increase is driven mainly by image and OT processing. The machine measurement uses strict-box F0/OT annotation-simulated answers and is separate from the Mild greedy protocol. The retained synthetic matrix spans 2/3/5/7/10/14 cameras, 5/20/50/100 people per view, two load modes, ten seeds, and three system repeats; it uses self-consistent sets and 128-dimensional artificial unit vectors, so it reports operator-scale behavior.

Reviewer_ZH47OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DAWSON, a framework for calibrating multi-view unique counts using a small number of aggregate shared-person answers. Its EER strategy selects camera pairs based on expected residual reduction. Experiments cover synthetic scenes, WILDTRACK, multiple encoders, and noisy answers. EER improves over Largest-overlap under degraded proxies, but remains substantially worse than the zero-query Max baseline.

Critical strengths
  1. Calibration using aggregate count answers has some exploratory value.
  2. The experiments cover multiple encoders, degradation levels, and answer noise.
Critical weaknesses
  1. The Max baseline substantially outperforms the proposed method. In Tables 10 and 15, Clean MAE is 0.190 for Max versus 1.75 for DINOv3 EER. Under Strong degradation, it is 1.605 versus 37.06. The proposed method incurs additional query costs while producing much worse counts than a simple zero-query baseline.
  2. Even with perfect counts and pairwise overlaps, the estimator is exact only when everyone appears in the same number of cameras. Otherwise, it systematically underestimates the total and can even fall below a single camera’s count. The authors acknowledge this limitation but neither validate the assumption nor correct the bias. Additional pairwise queries cannot resolve it.
  3. Pairwise overlaps cannot generally determine the unique count. For example, views {a,b}, {a,c}, and {a,d} contain 4 people, while {a,b}, {a,c}, and {b,c} contain 3 peaple. Both cases have identical per-view counts and pairwise overlaps. Even querying every camera pair cannot distinguish them, and the paper does not resolve this fundamental ambiguity.
  4. EER optimizes a surrogate residual whose reduction does not necessarily reduce actual counting error. The appendix acknowledges that the error can increase. Without sufficient component ablations and analysis, the connection between this acquisition objective and reliable counting improvements remains unclear.
  5. The paper has serious formatting and organization problems, including an entirely blank page, substantial whitespace, and multiple experimental protocols that are difficult to follow. The critical Max baseline appears late in the appendix, with insufficient discussion in the main text of its substantially better performance.
Questions for authors

Please refer to weakness for detailed questions. My major concern is whether the research problem is well-posed and well-addressed. What is the well-defined estimand DAWSON targets, and can the authors bound the irreducible error or provide a bias-corrected formulation rather than acknowledging the limitation without addressing it ?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ZH47 (Part 1/3 for Comment 1 and 2)

Thank you for your careful review and constructive comments. We have revised the manuscript to clarify the counting target, address the estimator’s limitations, strengthen the comparisons and analyses, and improve the presentation. We respond to each point below.

1.Comment: “The Max baseline substantially outperforms the proposed method. In Tables 10 and 15, Clean MAE is 0.190 for Max versus 1.75 for DINOv3 EER. Under Strong degradation, it is 1.605 versus 37.06. The proposed method incurs additional query costs while producing much worse counts than a simple zero-query baseline.“

1.Reply: In the submitted manuscript, same-input Max was already reported as a zero-query reference in the appendix. We now bring it into the main comparison tables and discuss its query cost and accuracy more directly. In the seven-view WILDTRACK evaluation, Max is substantially more accurate than two-query EER on the same windows and targets. The table also shows the same-budget Largest selector and a preselected view subset. MAE is lower when better; Max uses zero queries, while Largest and EER use two. | Setting | Condition | Max (0 queries) | Largest (2 queries) | EER (2 queries) | | full7 | Clean | 0.190 | 1.67 | 1.75 | | full7 | Strong | 1.605 | 51.74 | 37.06 | | C4/C5/C7 | Mild | 4.004 | 1.597 | 1.728 | In full7 Clean, Max equals the visible union in 84.0% of held-out windows. This annotation-based coverage statistic indicates that one view often already captures everyone in the selected-view union, explaining Max’s strong performance in this setting. Each subset has its own retained-visible union target, which changes with Mild degradation. The results use accurate annotation-simulated answers in one WILDTRACK scene . Section 4 and the main tables now include Max and query budgets.

2.Comment: “Even with perfect counts and pairwise overlaps, the estimator is exact only when everyone appears in the same number of cameras. Otherwise, it systematically underestimates the total and can even fall below a single camera’s count. The authors acknowledge this limitation but neither validate the assumption nor correct the bias. Additional pairwise queries cannot resolve it.”

2.Reply: The submitted method section already noted that heterogeneous visibility multiplicity makes the estimator approximate. We now characterize the resulting underestimation and validate the multiplicity assumption. For each window, the target is the union of people visible in the selected views; in the degradation study it is the retained-visible union. With exact view counts and all exact pair intersections in a nonempty window, the original estimator is N/(1+r), where r is the population variance of visibility multiplicity divided by its squared mean. It is exact for constant multiplicity and otherwise underestimates the union, potentially falling below a view count. Pair answers update overlap proxies but cannot generally remove this structural error. In the annotated retained sets, multiplicity is heterogeneous in 6,599 of 6,600 setting–seed–window records from one WILDTRACK scene. The formula links greater relative variation in visibility multiplicity to stronger underestimation. In the offline all-pair diagnostic, full7 Clean MAE is 1.863 before projection to [Max, Sum] and 0.186 afterward, versus 0.190 for zero-query Max. For C4/C5/C7 Clean, the corresponding all-exact MAEs are 2.007 and 1.875: the count constraint reduces error while leaving a substantial residual error. When the target lies in [Max, Sum], projection cannot increase its per-window absolute error. This is a separate post-processing diagnostic; the default output remains unclipped. The all-exact reference uses 3 or 21 pair intersections, whereas EER and Largest use two accurate simulated answers. In C4/C5/C7 Mild, all-exact MAE is 2.044 versus 1.597 for Largest and 1.728 for EER.

AuthorsOpenReview ↗

Response to Reviewer ZH47 (Part 2/3 for Comment 3)

3.Comment: “Pairwise overlaps cannot generally determine the unique count. For example, views {a,b}, {a,c}, and {a,d} contain 4 people, while {a,b}, {a,c}, and {b,c} contain 3 people. Both cases have identical per-view counts and pairwise overlaps. Even querying every camera pair cannot distinguish them, and the paper does not resolve this fundamental ambiguity.”

3.Reply: We agree that pairwise intersections do not generally identify the union. The reviewer’s examples both have view counts (2, 2, 2) and pair intersections (1, 1, 1), but unions of 3 and 4. We characterize the attainable interval [L, U] using nonnegative integer membership-pattern counts consistent with exact view counts and the revealed exact pair answers. The endpoints are attainable for these included constraints.We have added a new type of interval analysis, using it to represent the possible answers that still exist under the current information.For your example, the feasible interval is [3,4].

Adding an exact answer along the same trajectory cannot widen the interval. Its half-width is the minimax worst-case absolute error for a deterministic real-valued estimate based on these constraints. When the target satisfies the constraints, projecting the original output onto the interval cannot increase its absolute error; this separate post-processing does not change query selection or the default output.

With two accurate annotation-simulated answers, C4/C5/C7 Clean raw/interval-projected/zero-query Max MAE is 2.960/1.236/3.865. Projection moves estimates outside [L, U] to the nearest endpoint and leaves estimates inside unchanged.For example, if the feasible interval is [3, 4], an estimate of 5 is projected to 4, while an estimate already within [3, 4] remains unchanged.its benefit therefore depends on where the original estimate falls relative to the feasible interval. In full7 Mild, raw and projected EER both have MAE 12.764 versus 0.667 for Max. These are results from one WILDTRACK scene; the appendix also reports Largest and the offline all-pair reference.

AuthorsOpenReview ↗

Response to Reviewer ZH47 (Part 3/3 for Comment 4 and 5)

4.Comment: “EER optimizes a surrogate residual whose reduction does not necessarily reduce actual counting error. The appendix acknowledges that the error can increase. Without sufficient component ablations and analysis, the connection between this acquisition objective and reliable counting improvements remains unclear.”

4.Reply: The submitted manuscript defined EER as the expected decrease in an observable calibration residual. We now clarify its relationship to counting error: EER ranks queries by expected proxy-residual reduction under its fixed response distribution. The realized residual change after an answer and the offline change in absolute counting error are distinct. In particular, the revised manuscript now states: “EER optimizes the first quantity; neither realized residuals nor counting errors are constrained to decrease at every step.” We added trajectory and component analyses using DINOv3 greedy proxies, accurate annotation-simulated answers, and the retained-visible union in one WILDTRACK scene. In full7 Clean, MAE along the saved two-query trajectory is 1.598 → 1.723 → 1.747. In 185 of 400 query events, residual decreases while absolute error increases; these are two events per window across 200 windows. The residual measures observable consistency and unresolved proxy uncertainty, while actual error depends on the unknown union; reducing the former can therefore leave the estimate farther from the target. Across six settings, realized residual increases in 28 of 13,200 events. In C4/C5/C7 Mild, MAE is 3.094 → 1.583 → 1.728. The intermediate state belongs to the same two-query trajectory. Full/no-tightening MAE is 1.728/1.939 for C4/C5/C7 Mild and 12.764/10.829 for full7 Mild. Dynamic/fixed-initial ranking gives 1.728/1.560 and 12.764/12.855, respectively; fixed ranking still updates the estimate after each answer. These reversals show that the effect of each component depends on the evaluated view and degradation setting, making the ablations informative about where each design contributes.

5.Comment: “The paper has serious formatting and organization problems, including an entirely blank page, substantial whitespace, and multiple experimental protocols that are difficult to follow. The critical Max baseline appears late in the appendix, with insufficient discussion in the main text of its substantially better performance.”

5.Reply: The submitted manuscript described the individual experimental protocols and reported the zero-query Max comparison in the appendix. We reorganized the experimental presentation around the target, inputs, answer source, query budget and repetition unit. Section 4 points to an appendix protocol guide distinguishing synthetic scenes, WILDTRACK controls and degraded bridges, selected-view diagnostics, adapted representations and the offline human-answer pilot. We revised appendix page breaks and table placement. The main Clean and Clean/Mild tables include zero-query Max, and Section 4 reports full7 Clean Max/EER MAEs of 0.190/1.75 and Strong MAEs of 1.605/37.06, with zero and two queries, respectively. It also retains the subset results and same-budget Largest comparison. Some whitespace remains around long tables.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.