acceptodds
Under review as a conference paper at ICLR 2027

Retrospective Open-Vocabulary Memory for Long-Term Object Search

Abstract

Long-term object search requires learning where objects usually appear from repeated but uneven observations of a changing environment. We formulate retrospective open-vocabulary memory as probabilistic inference from censored observations, where the key idea is to reason with evidence per opportunity: a detection or non-detection should influence belief only in proportion to the robot's opportunity to observe the corresponding location. We introduce ECROM, which uses this principle to estimate long-term prevalence for concepts specified only at query time and converts the resulting belief directly into an active-search prior. To evaluate this problem, we introduce a controlled long-term benchmark in ten HM3D homes that independently varies object placement and observation opportunity across repeated traversals. ECROM improves support-level AP on held-out queries by 4.5 points and search SPL by 4.2 points over the strongest competing memory in each metric. The benchmark, dataset, and code will be open-sourced.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied AI, long-term memory, open-vocabulary object search, active perception, Bayesian inference, censored observations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pq8jOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduced a query-independent long-term visual memory based on the evidence per opportunity. It formulate retrospective open-vocabulary memory as censored inference over long-term object prevalence, explicitly separating semantic evidence from observation opportunity. And the paper proposed a benchmark called LOOP-Bench to test whether a robot can infer where an object usually occurs from repeated, uneven observations.

Critical strengths
  1. The modeling of long-term object occurrence under uneven observation is motivated. A robot deployed over weeks or months should learn where a object usually occurs. Distinguishing between an object being absent and the robot lacking an opportunity to observe it is a good point.

  2. LOOP-Bench provides a useful evaluation setting and helps study the effects of uneven observation. The real-robot experiments provide additional evidence of practical applicability, and the video in supplementary material helps illustrate the system’s behavior.

Critical weaknesses
  1. Retrospective querying currently incurs substantial latency. In the runtime analysis in appendix E, processing the stored 1,719 frames takes approximately 80 seconds per query on an RTX 5090. The appendix reports useful speed–accuracy trade-offs, but reducing the number of processed frames also reduces performance. Further analysis of how query latency scales with accumulated history would help assess its practicality during extended deployment.
Questions for authors
  1. Could authors extend the reported frame-reduction experiments to provide a more complete latency–performance curve and analyze query cost as the historical memory grows? It may also be useful to discuss whether frame selection or memory compression could keep latency manageable while preserving the benefits.
3 comments
AuthorsOpenReview ↗

Extended runtime analysis

We thank the reviewer for this suggestion. We ran the three analyses and report them in the Runtime part of Appendix E, with the new Table 4 and Figure 6, and the main text summarizes them in Section 6.5. In brief, scoring only the 20% of the keyframes most similar to the query reduces the query time from 80 s to 16 s at the same AP, the query time grows much more slowly than the history, and caching the image tokens of the detector during mapping answers a query in 13 ms with a similar AP and search success.

Latency-performance curve. We extended the frame-reduction experiment from two settings to shares between 5% and 100% of the stored keyframes, selected by SigLIP similarity to the query, and we evaluated estimation and search at each share. In the protocol of the submission, which recomputes the observation opportunity from the depth of the selected keyframes, AP is .191 at 50% of the keyframes and .182 at 20%, compared with .216 for exhaustive scoring. We additionally tested a setting that keeps the opportunity recorded during mapping instead of recomputing it. This opportunity is measured from the depth of all mapping frames, so it does not decrease when fewer keyframes are scored. At the same query time, this setting gives a higher AP: .217 at 20% of the keyframes (16 s instead of 80 s) and .203 at 5% (4.2 s).

Query cost as the memory grows. Query time is proportional to the number of stored keyframes, and this number grows much more slowly than the history, because CROSS stores a keyframe only when its view is not already matched by the map. The first traversal of a home stores 437 keyframes on average (49% of its frames), whereas the traversals after the tenth store about 12 each. Exhaustive query time is therefore 20 s after one traversal, 62 s after five and 80 s after twenty, and each of these later traversals adds about 0.6 s. AP saturates in the same way (.063, .190 and .216, Figure 6c).

Frame selection and memory compression. Frame selection by retrieval works as described above. We additionally test the idea to compress the memory before any query is known: keeping the 20% of the keyframes that cover the most support surface leaves AP at .217, with SR .540. The largest gain comes from caching. OWLv2 computes the class embedding and the box of each image token from the image alone, and the query enters only through its text embedding. We therefore run the image side once, when a keyframe is stored (44 ms), and answer queries from the stored tokens. Given the same text embedding, the cached logits are identical to those of the detector for all 713 million token-prompt pairs of the benchmark. A query then takes 28 ms with all tokens, and 13 ms when we keep the 256 most object-like tokens per keyframe. This budget reduces the token store from 14 GB to 0.7 GB per home, with the same AP and a similar search success (SR .567, compared with .560 for exhaustive scoring on the same GPU). With cached tokens, each further traversal adds less than a millisecond to a query.

SettingQuery timeAPAUROCSRSPL
Exhaustive detector (Table 1)80 s.216.760.587.381
Retrieval, top 50% of the keyframes, opportunity recomputed (submission)40 s.191.704.533.346
Retrieval, top 20%, opportunity recomputed16 s.182.651.447.311
Retrieval, top 5%, opportunity recomputed4.2 s.180.650.473.293
Retrieval, top 50%, opportunity from mapping40 s.221.750.553.366
Retrieval, top 20%, opportunity from mapping16 s.217.743.513.344
Retrieval, top 5%, opportunity from mapping4.2 s.203.726.573.354
Coverage pruning, 20% of the keyframes kept16 s.217.754.540.359
Cached tokens, all 5,184 per keyframe (14 GB per home)28 ms.214.759.567.370
Cached tokens, 256 per keyframe (0.7 GB)13 ms.214.759.567.370
Cached tokens, 64 per keyframe (173 MB)13 ms.215.760.547.373
Cached tokens, 16 per keyframe (43 MB)13 ms.221.759.560.378
Exhaustive detector on the GPU of the token cache80 s.214.759.560.363

Query times are for one query on an RTX 5090, averaged over the ten homes. The last row repeats exhaustive scoring on the GPU that computed the cached tokens, because bfloat16 detector scores differ in their last bits between GPU models. This rounding alone changes AP from .216 to .214 and SR from .587 to .560, so we regard SR differences of two to three points as within the variation of the search.

Thank you for the detailed response and additional analyses. The comparison between opportunity recorded during mapping and opportunity recomputed from selected keyframes is particularly interesting, as it directly concerns how observation opportunity is estimated. It is the central part of the paper.

Overall, I like the paper’s perspective and its focus on learning long-term object occurrence under uneven observation. The additional analyses address my concerns, and I am inclined to maintain my positive score.

Thank you for your encouraging feedback. We appreciate your thoughtful comments and are glad the additional analyses addressed your concerns.

Reviewer_k4nKOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ECROM addresses long-term open-vocabulary object search by estimating object prevalence from repeated RGB-D observations while accounting for uneven observation opportunities. It introduces a censored Bayesian inference model that combines calibrated detector evidence with geometric exposure and uses the resulting beliefs to guide active search. Experiments on the proposed LOOP-Bench and a physical robot demonstrate improvements over several memory baselines.

Critical strengths
  1. The paper identifies an important failure mode in long-term object search, i.e., confusing unobserved locations with locations where an object is absent.

  2. ECROM provides an interpretable probabilistic formulation that combines graded detector evidence with observation opportunity and includes targeted component ablations.

  3. LOOP-Bench offers a controlled evaluation of long-term object occurrence, complemented by experiments on a physical robot.

Critical weaknesses

The paper presents a well-motivated problem, a principled formulation, and a fairly complete experimental evaluation. I find no major weaknesses in the proposed approach. The following point is mainly intended to stimulate further discussion:

  1. Adaptation to evolving environments.

    The current formulation assumes a stationary object-placement distribution, which the authors have already acknowledged as a limitation. In real-world deployments, however, object-placement habits may gradually change over time. For example, an object that was frequently placed on a desk may later be more commonly placed in the kitchen. In such cases, earlier observations may continue to dominate the estimated prevalence, making it difficult to reflect recent changes. I would be interested to hear the authors' perspective on how ECROM could be extended to handle such evolving distributions. In particular, how might the framework distinguish genuine changes in object-placement habits from occasional fluctuations, while balancing long-term memory with adaptation to recent observations?

Questions for authors

Please see the weakness.

3 comments
AuthorsOpenReview ↗

Adaptation to evolving environments

We thank the reviewer for the positive assessment and for this question. We agree that habits which drift over time are the natural next step after the stationary case, and the revised Limitations (Section 8) name them as a direction for future research. Below we describe how ECROM would extend to them and what our current results suggest.

What ECROM does today. The posterior of Eq. 6 gives every traversal the same weight. After a habit change, the old location therefore keeps its probability mass until enough new evidence has accumulated, as the reviewer describes. The new Appendix I, summarized in Section 6.5, tests a milder departure from our assumptions, in which objects stay in place for several visits while their long-run frequencies are unchanged. ECROM keeps the highest search success in that setting, and the extension below addresses drift itself.

Adapting to recent observations. The posterior is a product of one factor per traversal, so recency can enter as a weight on each factor, which gives a recency-weighted posterior:

pγ(θj∣Hj)∝θjα0−1(1−θj)β0−1∏e=1E(1+θj,cje,(Λ(rje)−1))γE−e.p_\gamma(\theta_j \mid \mathcal{H}_j) \propto \theta_j^{\alpha_0-1}(1-\theta_j)^{\beta_0-1}\prod_{e=1}^{E}\Bigl(1+\theta_j\\,c_{je}\\,(\Lambda(r_{je})-1)\Bigr)^{\gamma^{E-e}}.

With γ=1\gamma = 1 this is ECROM, and γ\gamma below one limits the effective history to about 1/(1−γ)1/(1-\gamma) traversals. Old evidence then fades towards the prior, and the cost of a query is unchanged because the posterior remains a one-dimensional integral per support.

Genuine changes and occasional fluctuations. A fixed γ\gamma forgets at the same rate whether or not anything has changed. To forget only when the data require it, we would compare two explanations of the history of an object: a single prevalence for all traversals, or different prevalences before and after some traversal tt. Both marginal likelihoods follow from the same per-traversal factors, so their ratio can be computed for each query, for example with Bayesian online change-point detection and the likelihood of ECROM as its observation model. Two properties of the model help to separate a change from a fluctuation. First, an occasional absence is already expected under a prevalence below one. For example, an object with a prevalence of .3 on a desk is missing from it in 70% of the visits, so one miss changes this ratio little, whereas several well-observed misses at the old location together with detections at a new one change it strongly. Second, the evidence is weighted by observation opportunity, and a traversal that did not observe the desk contributes a factor of one. A change in the routes of the robot is therefore not mistaken for a change in the habits of the object, which is the same distinction that the paper makes in the stationary case.

Balancing long-term memory and adaptation. The search prior would then average the long-run posterior and the posterior after the change, weighted by the posterior probability that a change occurred. Without evidence of a change this is ECROM, and with strong evidence it uses only the traversals after the change.

What our results suggest. Adaptation has a price when nothing has changed, because a shorter history contains fewer placements. In Appendix I, the long-run estimate loses accuracy when a history holds fewer independent placements (AP .174, compared with .216). In the same appendix we derived the posterior for objects that stay (Eqs. 41 and 42), and this temporal model did not improve search over the unchanged posterior. We therefore expect the benefit of modeling change to depend on how well the recent traversals observe both the old and the new locations, and we would evaluate such an extension before adopting it. LOOP-Bench can provide this test: as in Appendix I, its histories can be recorded again with a placement distribution that changes during the history.

Reviewer_k4nKOpenReview ↗

Response by Reviewer k4nK

Thank you for the detailed and thoughtful response. I particularly appreciate the discussion of recency-weighted inference and Bayesian change-point detection. The distinction between genuine changes in object-placement habits and occasional fluctuations, while accounting for observation opportunities, is especially well considered. I also find the proposed balance between long-term memory and adaptation to recent observations reasonable. This satisfactorily addresses my question. Thank you again for the insightful discussion, and I wish you the best with the work!

Thank you for the thoughtful feedback. We’re glad that the discussion addressed your question, and we appreciate your careful reading and constructive comments.

Reviewer_a6mxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ECROM uses stored color and depth observations from repeated robot visits to estimate how often a queried object occurs near each candidate surface. Its Bayesian model combines detector scores with an estimate of how much of that surface the robot could see. The resulting probabilities guide where the robot searches. The evaluation uses ten simulated homes, 50 held-out queries and 150 search trials, plus 15 real-robot trials.

Critical strengths
  • The ability to answer new object queries from past observations is a useful contribution to long-term robot memory. Accounting for how much the robot could see also helps distinguish object frequency from viewing frequency, so an unseen surface is not treated as evidence of absence.
  • LOOP-Bench provides a well-controlled test of ECROM's ability to rank likely object locations by varying object placement separately from surface visibility. The methods share observations, detections, candidate surfaces and planner, and the count baseline also adjusts for observation coverage. ECROM improves average precision (AP) from .171 to .216, supported by a paired 95% confidence interval of [.005, .092] for the difference (Table 6).
Critical weaknesses
  • The evidence for long-term search remains limited because both simulated and physical protocols independently redraw object locations, leaving persistence and routines untested. The real-robot study is also small, covering one office and 15 trials, with 14/15 successes for ECROM and 11/15 for the baseline. Its changed detector calibration and updates to only one candidate surface per view without visibility weighting further limit how directly it validates the full simulated method.

  • The lack of a direct calibration test leaves the reliability of ECROM's occurrence probabilities unclear. The model treats the visible surface fraction as a proxy for the probability of seeing an object and uses the same detector-score mapping across viewing conditions. The sensitivity to view count and visible surface fraction in Table 5 makes it important to check whether predicted occurrence probabilities match actual frequencies.

  • The evidence for better search success or efficiency over the strongest baseline remains inconclusive, despite the supported improvement in location rankings. The success-rate gain of +.053 has a confidence interval of [−.013, .127], while the gain in success weighted by path length (SPL) is +.043 with an interval of [−.037, .117]. Both intervals include zero, so these results do not establish a reliable search advantage.

Questions for authors
  1. How does performance change if objects tend to remain in the same place for several visits, while their long-run location frequencies remain unchanged?
  2. For held-out objects, do predicted occurrence probabilities match actual frequencies across different numbers of views and fractions of each surface observed?
4 comments
AuthorsOpenReview ↗

Overview, persistence and routines

We thank the reviewer for the careful and constructive review. In response, the revision adds three experiments: persistent placements (Appendix I), a second set of real-robot experiments with the exposure-weighted belief update of our planner (Appendix G.1), and a direct evaluation of the predicted probabilities (Appendix H). The main text summarizes them in Sections 6.5 and 7. Table numbers below refer to the revised PDF, in which the tables after Table 3 of the submission shift by one (the paired intervals are now Table 7).

1. Persistence and routines (Weakness 1, Question 1)

We ran the experiment that the reviewer describes (Appendix I, Table 17, summarized in Section 6.5). Before each traversal, an object keeps its location with probability ρ\rho and is otherwise redrawn from its long-run distribution. The long-run frequencies and the ground truth are therefore unchanged, and ρ=0\rho = 0 is LOOP-Bench. With ρ=.5\rho = .5 and ρ=.8\rho = .8, an object stays for two and five traversals on average. The robot repeats the recorded actions, so the routes, keyframes, supports and opportunities are those of Table 1, and we search at the visit that follows the history (150 trials per condition). The table below gives the main rows of Table 17.

MethodAP / SR / SPL at ρ=0\rho = 0at ρ=.5\rho = .5at ρ=.8\rho = .8
ECROM (unchanged).216 / .587 / .381.200 / .593 / .367.174 / .500 / .360
ECROM with persistence model.216 / .587 / .381.191 / .573 / .355.163 / .487 / .345
PredictiveGraphs-style.160 / .533 / .339.134 / .527 / .334.120 / .433 / .319
DynaMem-style latest state.101 / .400 / .267.113 / .493 / .304.114 / .473 / .327
RAVEN-style retrieval.150 / .420 / .310.136 / .427 / .304.129 / .360 / .287
Detection count.148 / .440 / .330.128 / .433 / .321.114 / .400 / .319

We find three things. First, ECROM without any change has the highest search success at every ρ\rho. Its success rate is about the same at ρ=.5\rho = .5 and lower at ρ=.8\rho = .8, where the paired change from ρ=0\rho = 0 is −.087 [−.213, +.040]. To check whether persistence causes this decrease, we repeated the experiment with newly drawn placements, using three new histories per home. On these draws the success rates are lower and do not change with ρ\rho (.447, .433 and .460 for ECROM, paired change +.013 [−.107, +.140]), and ECROM again has the highest rate at each ρ\rho. The decrease in Table 17 is therefore likely specific to the history of the benchmark, about half of whose query placements are replaced at ρ=.8\rho = .8.

Second, the estimate of where an object usually is becomes less accurate (AP .174 at ρ=.8\rho = .8, compared with .216), because a history in which objects stay contains fewer independent placements. The other memories that accumulate detections also lose accuracy, and ECROM remains the most accurate.

Third, we derived the posterior for this process, which needs one additional parameter (Eqs. 41 and 42). It does not improve search for any ρ\rho, so we keep the model of the paper. Periodic routines are a natural next step, and the revised Limitations name them as a direction for future research.

AuthorsOpenReview ↗

Real-robot study and search advantage

2. Size of the real-robot study (Weakness 1)

We add additional 15 real-robot trials for each method. Now the revised paper reports two sets of real-robot experiments, 30 trials per method in total (Section 7, Table 2). The first set uses the detector-event update (Appendix G), and the second set uses the exposure-weighted belief update of our planner at three other placements per object (Appendix G.1). Both sets are scored with the same budget of 20 inspected views and 40 m of travel. Each cell gives the successful trials, the mean travel and the mean number of views in the successful trials.

ExperimentECROMDetection countPredictiveGraphs-style
Detector-event update (Table 12), 15 placements13/15, 19.3 m, 3.7 views9/15, 24.1 m, 4.6 views11/15, 27.2 m, 5.3 views
Exposure-weighted update (Table 13), 15 other placements14/15, 19.7 m, 3.7 views10/15, 26.8 m, 6.0 views9/15, 28.1 m, 6.7 views

Every placement is searched with all three memories from the same start, so we compare the memories paired by placement (Section 7). Over the 30 placements of the two sets, ECROM travels less than detection count and PredictiveGraphs-style memory, by 5.9 m and 8.2 m on average (Wilcoxon signed-rank test, p = .007 and p < .001). ECROM also travels less on the placements that both memories found (18 in each comparison): 17.1 m, compared with 20.6 m for detection count (p = .02), and 19.6 m, compared with 22.6 m for PredictiveGraphs-style memory (p = .04). It finds the target in 27 placements, compared with 19 and 20. ECROM succeeds where detection count fails in nine placements and the reverse happens in one (exact McNemar test, p = .02); against PredictiveGraphs-style memory the counts are nine and two (p = .07). The paper reports the two sets in separate columns because their belief updates differ. Both were run in the same office, and further environments are a natural next step.

3. Detector calibration and belief update on the robot (Weakness 1)

Belief update. The additional set uses the belief update of our search planner (Section 4.4). After every inspected view that does not end the search, Eq. 8, bj′∝bj,(1−pdet,cj(u))b'_j \propto b_j\\,(1 - p_{\mathrm{det}}\\,c_j(u)), lowers the belief of every support in proportion to its exposure cj(u)c_j(u) in that view, instead of updating one support per view without visibility weighting. The start, detector, candidate views and budget are those of the first set, and ECROM again succeeds most often (14 of 15 trials).

Likelihood ratio. Both settings fit the likelihood ratio with the same estimator, class histograms of the detector score for visible and not visible targets, made monotone and clipped (Appendix E.1), but from different labels. In LOOP-Bench, the simulator labels every episode-support cell of the calibration objects. On the robot, a VLM labels a score-stratified sample of stored keyframes as whole images, so that deployment in a new environment requires no human annotation, and we shrink these histograms towards those of the simulation. Both sets use this calibration (Table 11).

To measure the quality of these labels, we annotated 200 of the 1,080 VLM-labeled keyframes by hand, 100 drawn at random from each label. All 100 not-visible labels are correct, whereas only 48 of the 100 visible labels are, which means that the VLM over-reports visibility. Both sets use the likelihood ratio fitted on these noisy labels, and ECROM succeeds in 13 and 14 of their 15 trials. ECROM therefore transfers to the robot with noisy labels. We report this audit in Appendix G.

4. Search advantage over the strongest baseline (Weakness 3)

We agree that in simulation the paired intervals for success rate and SPL against PredictiveGraphs-style memory, the strongest baseline in search, include zero, and Section 6.3 states this. The revision adds evidence on search from the robot and from persistent placements. On the robot, over the 30 paired placements of item 2, ECROM travels 8.2 m less than PredictiveGraphs-style memory on average (p < .001) and finds the target in 27 placements, compared with 20. With persistent placements, ECROM has the highest success rate at every ρ\rho, on the histories of the benchmark and on newly drawn ones (item 1).

However, we emphasize that the primary contribution of ECROM is the exposure-calibrated estimation of long-term object prevalence, rather than a new navigation algorithm. On this central task, ECROM significantly outperforms the strongest search baseline, with paired improvements of +.056 AP [+.017, +.095], +.105 Top-5 mass [+.012, +.196] and +.101 AUROC [+.052, +.151]. Search is evaluated downstream of this estimate. All methods use the same search planner, so the differences in search are attributable to the memory.

AuthorsOpenReview ↗

Calibration of the occurrence probabilities

5. Calibration of the occurrence probabilities (Weakness 2, Question 2)

We added this test as Appendix H (Tables 14 to 16, Figure 7, summarized in Section 6.5), which compares the predicted prevalence of each of the 46,810 query-support pairs of the 50 held-out queries with its true frequency.

Overall. On aggregate, the predictions agree with the true frequencies: the ratio O/E of the mean true prevalence to the mean prediction is 1.03 [0.82, 1.24]. They are also close at the top of the ranking, where supports predicted above .1 hold the object 27% of the time for a mean prediction of 22%. Between 1% and 10%, however, the predictions are lower than the true frequencies, and these supports hold the object 3.6 to 3.9 times as often as predicted (logistic slope 1.30 [1.06, 1.60]). We traced this mainly to the prior, which we select for ranking. The prior that minimizes the NLL of the calibration objects brings the slope to 0.87, but it inflates the highest predictions and lowers AP from .216 to .210. We therefore keep the ranking prior and report this trade-off. ECROM also has the smallest systematic error of the compared memories (7.6% of its NLL, compared with 10.4% for PredictiveGraphs-style memory; Table 14).

By number of views and fraction observed (Question 2). Table 16 splits the supports by three observation conditions.

Observation conditionRangeO/E of ECROMO/E without exposure term
Traversals that observed the support1 to 31.16 [0.84, 1.48]3.17
4 to 61.07 [0.62, 1.56]2.52
7 to 200.85 [0.48, 1.22]1.35
Mean fraction of the surface observedbelow .81.46 [1.11, 1.83]3.20
.8 to below 10.80 [0.54, 1.13]1.66
10.82 [0.51, 1.13]1.95
Frames per traversal that observed itup to 110.96 [0.76, 1.15]2.19
above 11, up to 191.15 [0.80, 1.54]2.45
above 191.06 [0.59, 1.54]2.32

Across view-count strata, that is, the number of traversals and of frames per traversal that observed a support, aggregate predicted and observed prevalences agree within uncertainty: O/E stays between 0.85 and 1.16, and every interval includes one. Without the exposure term, O/E is between 1.35 and 3.17 in these groups, and the underestimation grows as a support is observed in fewer traversals.

Across the fraction of the surface observed, the agreement is weaker. On the third of the supports that is least completely observed, the object occurs 1.46 times as often as ECROM predicts, whereas the other two thirds agree with the predictions within their intervals. We traced this difference to the prior and not to the exposure term. The ratio of O/E between this third and the other supports is 1.80 [1.24, 2.76], it is the same without the exposure term (1.78), and it also appears when we simulate presence, visibility and detector responses from the model itself (1.86). In LOOP-Bench, objects are placed about twice as often on these partially observed supports, for the calibration objects as well, while all supports share one prior and 94% of them never receive positive evidence. When we give the two groups prior means in the ratio measured on the calibration objects, O/E becomes 1.02 and 1.05, and AP rises from .216 to .235 (paired difference +.019 [+.007, +.033]). The main results keep the single prior, and evaluating the new prior in search is a direction for future work.

The two approximations. A present object is visible in 3% of the traversals at zero exposure and in 87% at full exposure (Figure 7b), and replacing the exposure by the visibility curve measured on the calibration objects changes the probabilities little (systematic error 7.6% to 7.4%). For the shared detector-score mapping, aggregate predicted and observed prevalences agree within uncertainty in each view-count range of the sensor audit (O/E 0.96 to 1.15; Table 16).

We would appreciate your feedback on whether our additional experiments and analyses have addressed your concerns. If so, we hope you might consider revisiting your score in light of the new evidence. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.