acceptodds
Under review as a conference paper at ICLR 2027

Towards Active Cross-View Object Geo-Localization

Abstract

Cross-view object geo-localization (CVOGL) typically assumes a fixed query image, overlooking the ability of mobile agents to actively acquire more informative observations. To address this limitation, we introduce Active Cross-View Object Geo-Localization (ActiveGeo), where an agent sequentially selects new viewpoints and determines when to stop, aiming to improve localization with minimal observations. We further propose ActiveMoPT, an ActiveGeo framework with three-stage training. First, Multi-View Prompt-Preserving Adaptation enables the model to aggregate multiple query views while reusing the initial prompt. Second, Trajectory-Guided Policy Initialization uses supervised agent trajectories to learn viewpoint selection and initial stopping behavior. Third, Cost-Aware Policy Refinement employs GRPO with a gain-cost reward to jointly optimize localization accuracy and observation efficiency. We also construct ActiveGeo-858, a zero-shot test set containing 858 scenes and 1,716 target annotations. Experiments show that ActiveMoPT achieves state-of-the-art performance on MoP-UAV using only 1.45 query views on average, and substantially outperforms previous CVOGL approaches under zero-shot evaluation on ActiveGeo-858.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Active Cross-view object geo-localization, Post-Training, Localization Agent

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_hDjCOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces ActiveGeo, where an agent sequentially selects new viewpoints and determines when to stop. It proposes ActiveMoPT with three-stage training: Multi-View Prompt-Preserving Adaptation, Trajectory-Guided Policy Initialization, and Cost-Aware Policy Refinement.

Critical strengths
  1. ActiveGeo introduces a new task formulation that brings active perception to CVOGL through viewpoint selection and stopping.
  2. ActiveMoPT integrates a multi-view localizer and an action policy while reusing the initial prompt across subsequent unprompted UAV views.
  3. Experiments demonstrate a promising trade-off between localization accuracy and observation efficiency, including zero-shot evaluation on ActiveGeo-858.
Critical weaknesses
  1. The experiments rely on 54 predefined UAV viewpoints, with all corresponding images collected in advance. The agent selects among these viewpoints rather than actively navigating in an online environment. However, in practical UAV applications, the target location is initially unknown, raising the question of how these candidate viewpoints can be determined and accessed before the target has been localized.
  2. The action policy does not consider spatial dependencies between viewpoints. The agent can directly select any unobserved viewpoint without considering the physical movement required to reach it. In practice, moving to a selected viewpoint may involve passing through intermediate locations where useful observations could already be obtained. Moreover, the proposed reward assigns a fixed acquisition cost to each additional view, regardless of the travel distance or movement cost.
  3. The current experiments mainly compare ActiveMoPT with passive localization methods and random viewpoint acquisition. Comparisons with simpler selection strategies, such as fixed viewpoint priorities or heuristic-based selection, would help establish the advantages of the learned policy.
  4. A central contribution of ActiveMoPT is its ability to select informative viewpoints based on the current localization state. However, Table 8 reveals substantial preferences for specific viewpoint IDs, with View 54 accounting for 30.44% of second-view selections. Consequently, the evidence supporting the learned selection policy remains incomplete.
  5. Several implementation details make the reported improvements difficult to interpret precisely. First, Figure 2 describes the initial localizer as identical to MoPT, despite different single-view results in Table 2, leaving the extent of parameter adaptation unclear. Second, the Preserve-Improve loss is presented without weights in Eq. (6), whereas Appendix C.1 specifies nonuniform default weights. These ambiguities affect the reproducibility of the experiments and the interpretation of the proposed adaptation strategy.
Questions for authors
  1. In practical UAV applications where the target location is initially unknown, how can the 54 candidate viewpoints be determined and accessed?
  2. How would the spatial relationships between viewpoints, including intermediate observations during UAV movement and viewpoint-dependent travel costs, affect the learned action policy and its accuracy-efficiency trade-off?
  3. Could the authors compare ActiveMoPT against a fixed-priority viewpoint selection strategy (e.g., following the most frequently selected viewpoints in Table 8) under the same multi-view localizer and stopping mechanism?
8 comments
AuthorsOpenReview ↗

Response by authors (1)

We thank the reviewer for recognizing the ActiveGeo formulation that brings active perception to CVOGL, the reuse of the initial target prompt across subsequent unprompted UAV views, and the promising accuracy–observation-efficiency trade-off, including zero-shot evaluation on ActiveGeo-858. We provide additional experiments and clarify practical deployment, simpler selection strategies, and implementation details below.

Weakness 1 & Question 1: Determining and Accessing Candidate Viewpoints When the Target Location Is Unknown.

Thanks for raising this deployment question. The 54 UAV views are defined around the building at the center of the satellite image, which is not necessarily the target specified by the prompt. Defining these candidate viewpoints therefore does not require knowing the prompted target's location. The building at the center of the satellite image is used to define the candidate viewpoints, while the prompt specifies the target to be localized.

We have revised the viewpoint description in Appendix A (page 14).

Our current policy follows a decision loop similar to graph-based VLN. In graph-based VLN, an agent selects the next navigation node using language instructions, current observations, and history, then continues deciding from observations obtained upon arrival (Bridging the Gap [1], CVPR 2022). Our policy similarly uses the target prompt, localization state, and observation history to select the next candidate viewpoint or stop. In the current evaluation, selecting a viewpoint reveals its pre-collected image. For UAV flight, this loop could be extended so that a VLA policy determines observation positions from online feedback and a flight controller executes the corresponding movements, as explained below.

Weakness 2 & Question 2 (Part 1): Intermediate Observations during UAV Movement.

Thanks for highlighting useful observations available at intermediate positions. A VLA-style extension could use these observations to update localization, adjust movement, or stop before reaching the intended viewpoint. Our framework provides a foundation for anticipating which observations are useful and deciding when to stop.

VLA models map visual observations and language instructions to robot actions (OpenVLA [2], CoRL 2024). For UAV deployment, a VLA policy could take the target's language description and initial visual prompt, current onboard image, satellite reference image, localization history, and UAV state as inputs, and output relative movement and orientation commands. A flight controller could execute short movements from these commands, allowing the policy to decide again from newly acquired images. Observation positions could thus be determined and reached incrementally through online feedback.

For example, if a building partly blocks the target, the policy could anticipate that moving slightly left would reveal more. Images acquired during movement could then be used to update localization and decide whether to continue. If an intermediate observation already provides sufficient localization evidence, the policy could stop before reaching the intended position; if ambiguity remains, it could adjust its direction to seek a more informative view. Intermediate observations could therefore change both where the UAV observes and when it stops during movement.

The key decision throughout this loop is whether another observation is likely to improve localization enough to justify its cost. Our framework learns to make this decision through multi-view localization updates and policy training with the gain-cost reward, connecting the current localization state to viewpoint selection and stopping. These capabilities provide a foundation for extending active localization to continuous UAV control, as discussed in Appendix C.5. We examine viewpoint-dependent movement costs in Part 2 below.

We have revised the discussion in Appendix C.5 (page 19).

AuthorsOpenReview ↗

Response by authors (2)

Weakness 2 & Question 2 (Part 2): Viewpoint-Dependent Travel Costs and the Accuracy–Efficiency Trade-off.

Thanks for asking how viewpoint-dependent travel costs affect the learned policy. Without changing the architecture, training with a distance-based movement cost leads to more observations over a shorter estimated travel distance while maintaining localization accuracy. This demonstrates that the acquisition strategy adapts to the cost objective.

We keep the architecture, frozen localizer, and localization gain term unchanged, initialize the policy from SFT, and replace the fixed observation cost for CONTINUE with a movement cost. In this experiment, we compute movement cost using the normalized straight-line distance between consecutively selected viewpoints:

ct=λmoved(vt,vt+1)dref,c_t=\lambda_{\mathrm{move}}\frac{d(v_t,v_{t+1})}{d_{\mathrm{ref}}},

Here, λmove=0.010\lambda_{\mathrm{move}}=0.010 and dref=182.580 md_{\mathrm{ref}}=182.580\,\mathrm{m} is the mean distance between viewpoints in the training scenes. We retrain only the policy with GRPO and compare on the same 100 MoP-UAV test scenes and 17,447 targets using P+B+L. For the view-count row, accuracy and Avg. Views are from the paper; distance is estimated separately. We estimate travel distance by summing straight-line distances between consecutively selected viewpoints and report the average over targets:

CostmIoU ↑[email protected] ↑Avg. Views ↓Avg. Distance (m) ↓
View-count cost0.4140.4781.45075.761
Movement cost (based on distance)0.4160.4801.82248.077

The policy trained with movement cost has a shorter estimated travel distance, uses more observations, and maintains high localization accuracy. Reducing observation count and reducing movement distance are different objectives; the action policy changes with the cost definition. These results show that the same framework can train the action policy to account for movement costs while maintaining localization accuracy.

We have added the results and corresponding analysis to the main manuscript (Table 5, page 10).

AuthorsOpenReview ↗

Response by authors (3)

Weakness 3 & Question 3: Learned Viewpoint Selection Improves Localization over Simpler Strategies.

Thanks for suggesting comparisons with simpler selection strategies. With the same multi-view localizer and stopping mechanism, the learned policy achieves higher localization accuracy than frequency-weighted random selection and Fixed order at similar average budgets; these gains persist when observation counts are fixed. Below, we compare against two strategies favoring the frequently selected viewpoints in Table 12 (Table 8 in the original submission), then fix view counts to compare viewpoint selection strategies.

(1) With the same multi-view localizer and stopping mechanism, the learned policy achieves higher localization accuracy than both simpler strategies. To examine the reviewer's suggestion of favoring frequently selected viewpoints in Table 12, we use a frequency-weighted random baseline: sampling probabilities follow the table's selection frequencies, making frequently selected viewpoints more likely to be chosen, with the same fixed probabilities for all targets. We also include Fixed order, which follows a fixed priority ranking based on the learned policy's viewpoint-selection frequencies, selecting the highest-priority available, unobserved viewpoint at each step. All methods use the same initial view, multi-view localizer, candidate viewpoints, and GRPO stopping policy, changing only the viewpoint selection strategy. We use P+B+L on MoP-UAV and P+B on ActiveGeo-858.

DatasetViewpoint selection strategymIoU ↑[email protected] ↑Avg. Views ↓
MoP-UAVActiveMoPT (learned selection)0.41390.47781.4499
MoP-UAVFrequency-weighted random0.41360.47611.4478
MoP-UAVFixed order0.41300.47651.4486
ActiveGeo-858ActiveMoPT (learned selection)0.42220.47551.3706
ActiveGeo-858Frequency-weighted random0.42110.47141.3648
ActiveGeo-858Fixed order0.42140.47321.3648

Frequency-weighted random selection and Fixed order both inherit global viewpoint preferences from the learned policy's selection frequencies. Their competitive performance is consistent with these learned preferences being useful for localization. Both baselines apply the same global preferences across targets, whereas ActiveMoPT selects viewpoints using the current localization state and observation history. ActiveMoPT achieves additional improvements in mIoU and [email protected] on both datasets at similar average observation budgets, supporting the benefit of selecting viewpoints based on the current localization state.

We have added the results and corresponding analysis to the main manuscript (Table 4, page 9).

(2) Learned viewpoint selection improves localization at fixed view counts. Even with the same stopping policy and similar average budgets, different viewpoint choices can lead to different observation counts for individual targets. To isolate viewpoint selection more directly, we fix the observation count for each target. We group MoP-UAV targets by the number of views acquired before the policy stopped in the original paper evaluation. All viewpoint selection strategies are re-evaluated on the same targets, with identical initial observations and the same frozen localizer. For each subset, all acquire exactly K views, including the initial image, and then terminate. Only the viewpoint selection strategy differs.

Number of views KNumber of targetsLearned selection mIoU ↑Frequency-weighted random mIoU ↑Fixed order mIoU ↑
27,7210.40900.40830.4070
3550.24790.24130.2404
440.34780.34430.3414

With the same targets, initial observations, frozen localizer, and exactly K views per target, ActiveMoPT achieves higher localization accuracy than frequency-weighted random selection and Fixed order at each evaluated K. These differences reflect the selected viewpoints rather than differences in stopping behavior. These results support the policy’s ability to select more informative views for these targets based on the current localization state and observation history, even when the alternatives already use its learned global preferences.

We have added the results and corresponding analysis to Appendix B.2 (Table 8, page 15).

Reviewer_2ngPOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Active Cross-View Object Geo-Localization (ActiveGeo), where a UAV agent can acquire additional views and decide when to stop observing, rather than relying on a fixed query image. The authors propose ActiveMoPT, which extends MoPT through three training stages: multi-view adaptation to incorporate new observations while preserving the initial target prompt, supervised policy initialization using oracle trajectories, and GRPO-based refinement to balance localization accuracy and observation cost. The paper also introduces ActiveGeo-858, a zero-shot evaluation set with 858 scenes and 1,716 target annotations. Experiments on MoP-UAV and ActiveGeo-858 show that the proposed method improves localization performance over single-view baselines while using relatively few additional observations.

Critical strengths
  1. An interesting and well-motivated problem setting. I find the idea of extending cross-view object localization from passive prediction to active observation quite interesting. In practical UAV applications, the initial view may be occluded or unfavorable for matching, and acquiring another view can be more useful than repeatedly refining the same prediction. The paper formulates this problem clearly by considering both viewpoint selection and the decision of when to stop.

  2. A reasonable framework with encouraging results. The three-stage training pipeline is well organized, and the experiments provide useful evidence for the effectiveness of its components. I particularly appreciate that the authors evaluate observation efficiency rather than localization accuracy alone. The GRPO stage reduces the average number of views from 1.886 to 1.450 while slightly improving accuracy over SFT. The additional evaluation on ActiveGeo-858 is also valuable for assessing generalization beyond the original training scenes.

Critical weaknesses
  1. It is unclear how much of the improvement comes from genuinely adaptive viewpoint selection. My main concern is whether the learned policy actually selects views based on the current scene and localization state, or mainly exploits favorable viewpoints in the dataset. As described in Appendix A, the policy selects among 54 predefined viewpoint IDs arranged along three circular trajectories. Appendix Table 8 shows that View 54 alone accounts for 30.44% of second-view selections. This does not necessarily indicate a problem, but it raises the question of whether a much simpler viewpoint selection strategy could achieve similar results. The paper includes random-view ablations, but these do not provide a comparison under matched observation budgets. I would like to see a stronger baseline that always prioritizes the most effective fixed viewpoints, as well as a random policy evaluated at the same average number of views. Without these comparisons, I am not fully convinced that the learned policy provides a substantial advantage over simpler acquisition strategies.

  2. The active perception setting is still quite limited, and the benefits of policy refinement need clearer justification. The current framework selects from a predefined set of UAV images rather than planning actual camera movements. This is a reasonable simplification for an initial study, but it leaves an important practical question: how would these target-centered candidate viewpoints be determined when the target's geographic location is not yet known? The observation cost is also modeled as a fixed penalty per additional view, without considering movement distance or time. As a result, the current experiments mainly demonstrate efficient image selection rather than physically cost-aware UAV exploration. I also find the improvement from GRPO worth examining more carefully. In Table 2(b), SFT already improves mIoU from 0.392 to 0.411, whereas GRPO further increases it only to 0.414. Its main benefit appears to be reducing the number of observations. This is useful, but it would be more convincing to compare SFT and GRPO at similar observation budgets, rather than only reporting their default operating points.

Questions for authors

Q1. Could the authors compare the learned viewpoint selection policy against simpler alternatives under the same average observation budget? In particular, a fixed strategy that prioritizes the most frequently selected or most effective viewpoints would be helpful. Given the strong preference for View 54 in Appendix Table 8, this comparison seems important for understanding whether the policy truly adapts to different scenes.

Q2. How would the predefined viewpoint setting translate to a practical UAV scenario? The 54 candidate viewpoints are arranged around the target, but the target's geographic location is precisely what the system is trying to estimate. Could the authors clarify what information is assumed to be available for defining these candidate viewpoints, and how movement costs would affect the current formulation?

Q3. Could the authors provide an accuracy-versus-observation-cost comparison between SFT and GRPO? For example, evaluating both policies at similar average numbers of views would help clarify whether GRPO learns a better acquisition strategy or primarily changes the stopping behavior. This would make the contribution of the RL stage more convincing.

6 comments
AuthorsOpenReview ↗

Response by authors (1)

We thank the reviewer for recognizing the interesting and well-motivated problem setting, the well-organized three-stage training pipeline, and the value of evaluating observation efficiency and generalization on ActiveGeo-858. We provide additional experiments and clarifications on adaptive viewpoint selection, practical deployment, and the benefits of GRPO below.

Weakness 1 & Question 1: Independent Contribution of Learned Viewpoint Selection.

Thanks for raising this concern. The learned policy improves accuracy at similar budgets, benefits multi-step localization at identical observation counts, and uses the current scene and localization state to assess the next view’s additional information and adapt its selection. Below, we present the comparison results and explain how the policy selects viewpoints based on the current state.

(1) Learned viewpoint selection achieves higher localization accuracy at similar average observation budgets. We compare the learned viewpoint selection policy against two simpler alternatives. Frequency-weighted random selection samples viewpoints according to their selection frequencies in Appendix Table 12 (Table 8 in the original submission). These probabilities are fixed across targets. Fixed order follows a fixed priority ranking based on the learned policy's viewpoint-selection frequencies, selecting the highest-priority available, unobserved viewpoint at each step. All methods use the same initial view, localizer, candidate viewpoints, and GRPO stopping policy. We use P+B+L on MoP-UAV and P+B on ActiveGeo-858.

DatasetViewpoint selection strategymIoU ↑[email protected] ↑Avg. Views ↓
MoP-UAVActiveMoPT (learned selection)0.41390.47781.4499
MoP-UAVFrequency-weighted random0.41360.47611.4478
MoP-UAVFixed order0.41300.47651.4486
ActiveGeo-858ActiveMoPT (learned selection)0.42220.47551.3706
ActiveGeo-858Frequency-weighted random0.42110.47141.3648
ActiveGeo-858Fixed order0.42140.47321.3648

Frequency-weighted random selection and Fixed order both inherit global viewpoint preferences from the learned policy's selection frequencies. Their competitive performance is consistent with these learned preferences being useful for localization. Both baselines apply the same global preferences across targets, whereas ActiveMoPT selects viewpoints using the current localization state and observation history. ActiveMoPT achieves additional improvements in mIoU and [email protected] on both datasets at similar average observation budgets, supporting the benefit of selecting viewpoints based on the current localization state.

We have added the results and corresponding analysis to the main manuscript (Table 4, page 9).

AuthorsOpenReview ↗

Response by authors (2)

(2) Learned viewpoint selection benefits multi-step localization at identical observation counts. Even with the same stopping policy and similar average budgets, different viewpoint choices can lead to different observation counts for individual targets. To isolate viewpoint selection more directly, we fix the observation count for each target. We group MoP-UAV targets by the number of views acquired before the policy stopped in the original paper evaluation. All methods are re-evaluated on the same targets, with identical initial observations and the same frozen localizer. For each subset, all acquire exactly K views, including the initial image, and then terminate. Only viewpoint selection differs.

Number of views KNumber of targetsLearned selection mIoU ↑Frequency-weighted random mIoU ↑Fixed order mIoU ↑
27,7210.40900.40830.4070
3550.24790.24130.2404
440.34780.34430.3414

With the same targets, initial observations, frozen localizer, and exactly K views per target, ActiveMoPT achieves higher localization accuracy than frequency-weighted random selection and Fixed order at each evaluated K. These differences reflect the selected viewpoints rather than differences in stopping behavior. These results support the policy’s ability to select more informative views for these targets based on the current localization state and observation history, even when the alternatives already use its learned global preferences.

We have added the results and corresponding analysis to Appendix B.2 (Table 8, page 15).

(3) The current scene and localization state help the policy assess the additional information offered by the next view. A favorable viewpoint in the dataset is useful for localization overall, but does not necessarily provide the most additional information at every step. The value of the next view depends on the current scene, previous observations, and current localization result. After each new observation, the policy uses the updated localization state and observation history to select the next viewpoint or stop (Fig. 2, Sec. 4.1).

View 54 often has an orientation and spatial layout closer to the satellite image, facilitating cross-view matching and making it the most frequent second-view choice. Yet it accounts for only 30.44% of second-view selections. Third-view choices are more dispersed, led by View 45 (22.03%), followed by Views 37 and 53 (Table 12, Appendix C.4). Thus, although the policy favors certain viewpoints, its choices are not concentrated on a single fixed viewpoint. These comparisons support the policy’s ability to use the current scene and localization state to select informative views and improve localization.

We have revised the analysis in Appendix C.4 (pages 18–19).

AuthorsOpenReview ↗

Response by authors (3)

Weakness 2 & Question 2: Candidate Viewpoint Assumptions and Movement Cost.

Thanks for raising this practical concern. The predefined viewpoints are associated with the scene and do not require the prompted target's location. A VLA-style extension could connect viewpoint selection with UAV movement and online observations. Incorporating movement costs into the reward changes the acquisition strategy, reducing estimated travel distance while maintaining localization accuracy.

(1) Defining the candidate viewpoints does not require knowing the prompted target's location. The 54 candidate viewpoints are arranged around the building at the center of the satellite image, which is not necessarily the target whose geographic location we aim to estimate. The benchmark provides this scene's predefined candidate viewpoints, while the prompt specifies the target to be localized. Defining these viewpoints therefore does not require knowing the prompted target's geographic location.

We have revised the viewpoint description in Appendix A (page 14).

(2) A VLA-style extension could connect viewpoint selection with UAV movement and online observations. Our current policy selects the next candidate viewpoint or stops based on the current state, similar to graph-based VLN. In graph-based VLN, an agent selects the next navigation node using language instructions, current observations, and history, then continues deciding from observations obtained upon arrival (Bridging the Gap [1], CVPR 2022). Our policy similarly uses the target prompt, localization state, and observation history to select an unobserved viewpoint or stop. The benchmark provides the corresponding pre-collected image after selection.

For UAV flight, we consider VLA-style control with continuous feedback, where observation positions could be determined incrementally from current observations. VLA models map visual observations and language instructions to robot actions (OpenVLA [2], CoRL 2024). Given the target's language description and initial visual prompt, current onboard image, satellite reference image, localization history, and UAV state, a VLA policy could output relative movement and orientation commands for a flight controller to execute. For example, suppose a building partly blocks the target. The policy could anticipate that moving slightly left would reveal more, make a short movement, then use the new image to decide whether to adjust position further, update localization, or end observation and return the result. Observation positions would emerge through this loop without requiring the target's precise coordinates in advance.

Regardless of the underlying controller, anticipating the localization value of subsequent observations from the current state is a key capability for active localization; our work provides a foundation for learning it. Before acquiring another observation, the policy must assess which viewing direction could supply useful cues and whether the expected benefit justifies further cost. Our framework combines multi-view localization with policy training using the gain-cost reward to learn observation selection and stopping with discrete viewpoints, providing a foundation for extending these decisions to continuous UAV control. We discuss this extension in Appendix C.5.

We have revised the discussion in Appendix C.5 (page 19).

Reviewer_YAQyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper studies active observation for Cross-View Object Geo-Localization (CVOGL). Existing methods typically rely on fixed UAV query images for object localization and cannot actively acquire additional viewpoint information. To address this limitation, the authors propose ActiveGeo, which extends conventional passive localization into a sequential decision-making task where an agent can autonomously select viewpoints and decide when to stop observing.

The authors further propose ActiveMoPT, which consists of three main stages:

1,Multi-view prompt-preserving adaptation: The method preserves the initial object prompt while progressively incorporating new UAV images through a multi-view attention mechanism. A Preserve-Improve Loss is introduced to reduce the negative impact of additional viewpoints on previous localization results.

2.Trajectory-based policy initialization: Expert trajectories are constructed in advance and used for supervised fine-tuning to learn viewpoint selection and stopping decisions.

3.Cost-aware policy optimization: GRPO and a gain-cost reward mechanism are employed to further optimize the viewpoint selection policy, balancing localization accuracy and the number of observations.

In addition, the authors construct the ActiveGeo-858 dataset to evaluate the model's zero-shot generalization to unseen scenes. Experimental results show that ActiveMoPT achieves promising localization performance on both MoP-UAV and ActiveGeo-858, while reducing the average number of observations.

Critical strengths

S1. The research problem is interesting, and the task formulation has practical significance.

This paper extends conventional cross-view object localization with fixed viewpoints into an active observation task. The UAV can autonomously acquire new viewpoints based on its current localization state and decide when to stop. This design considers the need for active decision-making in practical mobile agent scenarios. In addition, by preserving the initial prompt information, the method allows subsequent images without new prompts to continue localizing the same target. This is a meaningful aspect of the method design.

S2. The experimental evaluation is relatively comprehensive and includes generalization tests on independent scenes.

The authors conduct experiments on two datasets, MoP-UAV and ActiveGeo-858, and compare localization performance under different prompt configurations. ActiveMoPT consistently outperforms the original MoPT under several spatial-prompt settings. For example, on MoP-UAV with the P+B setting, mIoU increases from 0.375 to 0.395, while on ActiveGeo-858, it improves from 0.394 to 0.422. Furthermore, the ablation studies show that GRPO further improves localization performance while reducing the average number of observations from 1.886 to 1.450.

Critical weaknesses

W1. The independent contributions of active viewpoint selection and stopping decisions have not been sufficiently validated.

This paper jointly trains viewpoint selection and stopping decisions within a policy network. However, the current experiments do not sufficiently distinguish the individual contributions of these two components.

In Table 2, the random multi-view strategy uses an average of 2.5 viewpoints, whereas the complete active policy uses only 1.45. Since the observation budgets are different, it remains unclear whether the performance improvement mainly comes from more informative viewpoint selection or from the reduced number of observations enabled by the adaptive stopping mechanism.

Additional comparisons using the same stopping policy and similar observation budgets would help verify the contribution of active viewpoint selection itself.

W2. The definition of observation cost is relatively simple and does not fully reflect actual UAV movement costs.

This paper mainly measures efficiency by the number of additional observations and assigns a fixed cost to each additional observation in the reward function. However, in practical UAV applications, the movement distance, flight time, and energy consumption between different viewpoints can vary significantly. Simply counting the number of observations cannot fully capture the actual movement costs.

Therefore, the current experiments mainly demonstrate efficiency improvements in terms of observation count, while the benefits in terms of actual UAV movement efficiency remain insufficiently validated.

W3. The experimental results lack an analysis of statistical stability.

Although ActiveMoPT achieves promising results on both datasets, the paper mainly reports results from individual experiments without providing performance variations across different random seeds or confidence intervals.

This is particularly relevant to the GRPO stage, where mIoU only increases from 0.411 to 0.414, representing a relatively small improvement. Although the average number of observations decreases more noticeably, it is still difficult to determine whether the localization improvement is consistent and stable.

Additional experiments with multiple independent runs and statistical analysis would further strengthen the reliability of the experimental conclusions.

Questions for authors

Q1. Could the authors further distinguish the individual contributions of active viewpoint selection and adaptive stopping? For example, under the same stopping policy and a similar average number of observations, could the authors provide comparisons where the learned viewpoint selection strategy is replaced with random selection or a fixed viewpoint priority?

Q2. This paper treats each additional observation as having a fixed cost. However, in practical UAV operations, movement distances and energy consumption may vary across different viewpoints. Have the authors considered incorporating actual movement costs into the reward function? Would doing so affect the viewpoint selection behavior of the current policy?

Q3. The mIoU improvement introduced by the GRPO stage is relatively small. Could the authors provide experimental results across different random seeds, together with standard deviations, to verify whether the performance improvements from policy optimization are stable?

4 comments
AuthorsOpenReview ↗

Response by authors (1)

We thank the reviewer for recognizing the meaningful extension from passive localization to active viewpoint acquisition while preserving the original target prompt, and the encouraging improvements in localization accuracy and observation efficiency. We provide additional controls to isolate viewpoint selection from stopping and examine movement costs below.

Weakness 1 & Question 1: Independent Contribution of Viewpoint Selection.

Thanks for raising this concern. To disentangle viewpoint selection from stopping behavior, we compare learned and random selection under the same stopping policy and then at fixed observation counts for each target. Learned viewpoint selection improves localization in both experiments, supporting the independent contribution of selecting informative views.

(1) Learned viewpoint selection achieves higher localization metrics at similar average observation budgets. Within each dataset, we keep the initial observations, prompts, frozen multi-view localizer, candidate viewpoints, and GRPO stopping policy unchanged, changing only the viewpoint selection strategy. Frequency-weighted random selection uses a fixed distribution derived from the learned policy's empirical selection frequencies, corresponding to the preferences summarized in Appendix Table 12 (Table 8 in the original submission). We use P+B+L on MoP-UAV and P+B on ActiveGeo-858.

DatasetViewpoint selection strategymIoU ↑[email protected] ↑Avg. Views ↓
MoP-UAVActiveMoPT (learned selection)0.41390.47781.4499
MoP-UAVFrequency-weighted random0.41360.47611.4478
ActiveGeo-858ActiveMoPT (learned selection)0.42220.47551.3706
ActiveGeo-858Frequency-weighted random0.42110.47141.3648

Frequency-weighted random selection inherits global viewpoint preferences from the learned policy's selection frequencies. Its competitive performance is consistent with these learned preferences being useful for localization. This baseline applies the same global preferences across targets, whereas ActiveMoPT selects viewpoints using the current localization state and observation history. ActiveMoPT achieves additional improvements in mIoU and [email protected] on both datasets at similar average observation budgets, supporting the benefit of selecting viewpoints based on the current localization state.

We have added the results and corresponding analysis to the main manuscript (Table 4, page 9).

(2) Learned selection improves multi-step localization at fixed view counts. Even with the same stopping policy and similar average budgets, different viewpoint choices can lead to different observation counts for individual targets. To isolate viewpoint selection more directly, we fix the observation count for each target. We group MoP-UAV targets by the number of views acquired before the policy stopped in the original paper evaluation. Both viewpoint selection strategies are re-evaluated on the same targets, with identical initial observations and the same frozen localizer. For each subset, both acquire exactly K views, including the initial image, and then terminate. Only viewpoint selection differs.

Number of views KNumber of targetsLearned selection mIoU ↑Frequency-weighted random mIoU ↑
27,7210.40900.4083
3550.24790.2413
440.34780.3443

With the same targets, initial observations, frozen localizer, and exactly K views per target, ActiveMoPT achieves higher localization accuracy than frequency-weighted random selection at each evaluated K. These differences reflect the selected viewpoints rather than differences in stopping behavior. These results support the policy’s ability to select more informative views for these targets based on the current localization state and observation history, even when the baseline already uses its learned global preferences.

We have added the results and corresponding analysis to Appendix B.2 (Table 8, page 15).

AuthorsOpenReview ↗

Response by authors (2)

Weakness 2 & Question 2: Movement Costs and Observation Efficiency.

Thanks for raising this practical concern. To examine how movement costs affect acquisition behavior, we replace the observation-count cost with a movement cost while keeping the architecture and localizer unchanged. The policy reduces the estimated travel distance while maintaining localization accuracy and using more observations. This shows that the same ActiveMoPT framework can adapt its acquisition strategy to different cost objectives.

(1) Accounting for movement cost leads to more observations over a shorter estimated travel distance while maintaining localization accuracy. We keep the architecture, frozen localizer, and localization gain term unchanged, initialize the policy from SFT, and replace the fixed observation cost for CONTINUE with a movement cost. In this experiment, we compute movement cost using the normalized straight-line distance between consecutively selected viewpoints:

ct=λmoved(vt,vt+1)dref,c_t=\lambda_{\mathrm{move}}\frac{d(v_t,v_{t+1})}{d_{\mathrm{ref}}},

Here, λmove=0.010\lambda_{\mathrm{move}}=0.010 and dref=182.580 md_{\mathrm{ref}}=182.580\,\mathrm{m} is the mean distance between viewpoints in the training scenes. We retrain only the policy with GRPO and compare on the same 100 MoP-UAV test scenes and 17,447 targets. We estimate travel distance by summing straight-line distances between consecutively selected viewpoints and report the average over targets:

CostmIoU ↑[email protected] ↑Avg. Views ↓Avg. Distance (m) ↓
View-count cost0.4140.4781.45075.761
Movement cost (based on distance)0.4160.4801.82248.077

The policy trained with movement cost has a shorter estimated travel distance, uses more observations, and maintains high localization accuracy. Reducing observation count and reducing movement distance are different objectives; the acquisition policy changes with the cost definition.

(2) The results with movement cost reinforce the contribution of our framework. Our core contribution is an active localization framework that integrates new observations, selects informative views, and decides when to stop. Observation count defines one cost objective within this framework. This experiment demonstrates that ActiveMoPT learns an effective acquisition strategy under a different objective without changing the architecture. This adaptability strengthens the framework's relevance to practical UAV localization and establishes an empirical foundation for training policies with movement costs tailored to deployment.

We have added the results and corresponding analysis to the main manuscript (Table 5, page 10).

I have read your response, and it has addressed my concerns. I will increase my rating accordingly.

AuthorsOpenReview ↗

Follow-up Response by authors

Thank you for raising your rating and confirming that our previous responses have addressed your concerns. We appreciate your recognition of the additional experiments and clarifications.

We noticed that there is a newly added question about the stability of RL gains across random seeds. Given the limited time remaining, we could not conduct a new set of training runs across multiple seeds. We therefore reviewed the experiments already completed during the rebuttal period and consolidated the available results below.

Specifically, our experiments addressing your earlier question about movement cost included three independent GRPO training runs for each movement cost coefficient, using seeds 11, 22, and 33. All runs start from the same SFT checkpoint and keep the localizer frozen. We evaluate them on the same 100 test scenes and 17,447 targets. The table reports the shared SFT baseline and the mean ± sample standard deviation across the three training seeds for each GRPO setting:

Method / Movement cost coefficientmIoU ↑[email protected] ↑Avg. ViewsEstimated Geometric Distance (m) ↓
SFT (shared initialization)0.4114310.4736061.8857113.418
GRPO, λ = 0.0010.417137 ± 0.0006500.478287 ± 0.0008451.8409 ± 0.0245101.065 ± 3.971
GRPO, λ = 0.0030.416276 ± 0.0004160.477484 ± 0.0004871.8515 ± 0.009779.101 ± 3.044
GRPO, λ = 0.0100.416371 ± 0.0000780.478955 ± 0.0006841.7902 ± 0.028645.627 ± 2.522
GRPO, λ = 0.0300.414578 ± 0.0006710.475497 ± 0.0015481.5148 ± 0.084322.948 ± 3.727
GRPO, λ = 0.1000.405472 ± 0.0000050.463232 ± 0.0000001.0000 ± 0.00000.001 ± 0.002

At λ = 0.001, 0.003, 0.010, and 0.030, all three training seeds achieve higher mIoU than the shared SFT baseline. In particular, at λ = 0.010, the setting used in our previous response, the mean mIoU improvement is 0.004940, with a standard deviation across seeds of 0.000078. Both the average observation count and estimated geometric distance also decrease relative to SFT. These results provide direct evidence of consistent RL gains across training seeds under the movement cost setting.

The complete results also show how the cost objective affects the policy. As the movement cost coefficient increases, the policy generally favors shorter travel distances. At λ = 0.100, the high cost leads all three runs to stop almost immediately, reducing localization accuracy. This behavior is consistent with the localization gain–cost trade-off optimized by RL.

Thank you again for your careful review, constructive suggestions, and acknowledgment that our previous responses have addressed your original concerns. Your feedback has helped us present the contributions and experimental evidence more clearly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.