DiffVAS: Diffusion-Guided Visual Active Search in Partially Observable Environments

Thank you for reviewing our paper! We address all your concerns below.

Q1: There are writing issues such as inconsistent capitalization of symbols like "Sgn" in the formulas (6) and (7).

A1: Thank you for pointing this out. We have fixed it in the updated version, and will carefully go over the rest of the style and formats to be consistent everywhere.

Q2: Figure 3 has issues related to font, capitalization, and symbol representation. There is also a missing parenthesis in "Past Query Outcomes," which affects the clarity and professionalism of the presentation.

A2: Thank you again for catching this. We have fixed the missing parenthesis issue in the updated draft.

Q3: In the "Zero-shot Generalization" section, there is a misexpression where it states "solely on DOTA on xView".

A3: Thank you, we apologize for the oversight and the incorrect phrasing. The "on xview" part was included by mistake, and we have already fixed it in the updated version of the draft.

Q4: The paper’s ablation study on the reward function lacks the combination of $R^{AS}+R^{LU}$, which could provide deeper insights into the impact of these components on model performance. Why was the combination of $R^{AS}+R^{LU}$ omitted in the reward function ablation study?

A4: We fully agree with your point. Following your suggestion, we conducted an additional experiment where we trained DiffVAS using the $R^{AS} + R^{LU}$ reward function and compared its performance with other reward function variations. These additional experimental results have already been included in Appendix Section A11 of the updated draft.

We observe a slight improvement in search performance when incorporating $R^{LU}$ into the reward function for training the TCPM module, underscoring the importance of the local uncertainty-based reward factor ($R^{LU}$). Finally, note that the full reward is still the best, as seen by comparing it with Table 5 in the main paper.

Q5: Is the presence of issues in writing and formatting, such as text descriptions and figure/table formats, an indication that the paper was prepared somewhat hastily?

A5: Absolutely not! We sincerely apologize for the formatting and writing mistakes. Please rest assured that we will carefully address and resolve all such issues in the final version, to ensure a polished and well-presented paper.

Thanks, we appreciate the reviewer's valuable feedback! We've included all the additional details as requested.

Q1 Why "ANT" metric?

A1 Thank you for the suggestion. Exploration is crucial for solving this task, making it possible to optimize a policy for higher ANT at the expense of increased path length. Since our primary focus is maximizing target object discovery within a given budget, we evaluate performance using ANT. In scenarios where movement cost outweighs acquisition cost, path length would be a more suitable metric. However, in our work, we assume movement cost is negligible compared to acquisition cost (for example, acquisition cost might include sending a team on the ground for verification or having a human analyst review imagery), a premise we believe is practical in most cases.

Q2 Multi-scale AVS to simulate realistic search scenarios?

A2 To the best of our knowledge, this is the first attempt to address Visual Active Search in Partially Observable Environments. Our work is designed to serve as a foundational contribution to the emerging field of "Active Search in Partially Observable Environments." We believe this pioneering approach will pave the way for numerous future advancements in this important area. We are truly excited about the prospect of exploring this suggested direction in future work, as we agree that incorporating multi-scale AVS would bring our approach even closer to real-world deployability. Note, however, that there are also real-world scenarios where a UAV may have to hover at a fixed altitude -- for example, in a hurricane-hit area with strong winds, flying at a constant and quite low altitude may be necessary to avoid heavier winds higher up.

Q3 Future work could explore more challenging detection settings to evaluate DiffVAS’s resilience under realistic conditions.

A3 We would like to emphasize that DiffVAS is evaluated across a wide range of target object categories and diverse scenes, showcasing its robustness. Additionally, integrating an off-the-shelf object detection module could further enhance DiffVAS to address practical challenges like occlusions, noise, and other real-world obstacles. We agree with you that this could be explored in future work, as our current effort serves as a foundational contribution to the emerging field of "Active Search in Partially Observable Environments," with the potential to drive future advancements in handling more complex real-world detection challenges.

Q4 Computational overhead of DiffVAS? Approximate time required for the diffusion model to reconstruct a full image?

A4 We acknowledge the additional computational complexity introduced by the diffusion-based CGM module. However, our end-to-end DiffVAS framework infers the next region to query in approximately 2.20 seconds on a standard GPU. Specifically, a diffusion-based CGM module approximately takes 1.07 seconds to reconstruct a single full image on NVIDIA V100 GPU. Given that verifying a search query on the ground by a park ranger typically takes much longer (in the order of minutes to hours), the response time of our system is well within the operational requirements. This efficiency ensures that our framework can provide timely and actionable support, making it highly suitable for real-world deployment where swift decision-making is crucial. These additional details have been included in Appendix A9 of the updated draft.

Q5 (1) Further explanation about the cross-attention layer in Fig.3? (2) What's the reference and the target features? What's AdaIN?

A5 (1) We had already analyzed the suggested ablation study in our original submission. We've analyzed the efficacy of the cross-attention layer and reported our findings in Appendix A6, titled "Effectiveness of Cross-Attention Layer". Our findings indicate that the cross-attention layer is effective in learning target-aware representation.

(2) By "reference features," we refer to the latent features extracted by the CGM module, specifically the latent representation of the reconstructed image (i.e., $l_{img}(t)$). On the other hand, "target features" refer to the latent features of the target category, which are computed using the CLIP-based text encoder (i.e.,$l_{z}$). Additionally, "AdaIN" stands for "adaptive instance normalization," as originally proposed in [1]. We apologize for omitting these details. These additional details have been included in Appendix A12 of the updated draft.

Q6 Assessing the contribution of CGM?

A6 We previously performed the suggested ablation study in our original submission, and the details are provided in Appendix A7. The empirical results clearly demonstrate the significant advantage of the diffusion-based CGM module, which reconstructs the search area on the fly.

References:

1. Arbitrary Style Transfer in Real-Time With Adaptive Instance Normalization; Xun Huang et. al.; ICCV, 2017

We appreciate the reviewer's comments and address all the concerns below.

Q1: The idea of reconstructing the whole from parts has already been explored in many published works. The approach overlaps significantly with DiffMAE(ICCV23), which combines MAE with diffusion models for image reconstruction. The contribution seems incremental without clear novelty. The method closely resembles MDT(ICCV23), which uses unmasked tokens to predict masked ones while preserving diffusion training.

A1: Fundamental difference with prior work: The problem we are tackling in this work are fundamentally different than the mentioned work, such as MDT (ICCV23) and DiffMAE (ICCV23). Both these works aim to reconstruct an image from the partially observed scene, while we tackle a more challenging problem of active discovery of target objects. Hence, these methods (such as MDT, and DiffMAE) typically focus solely on optimizing for reconstruction, while our ultimate goal is identifying target-rich regions. Success for our task hinges on balancing exploration (obtaining useful information about the scene) and exploitation (finding objects of interest). To justify why such a reconstruction-only method is not appropriate to tackle active search problems, we performed an experiment with Greedy-DiffVAS that utilized the reconstructed features only to decide the next query location and compared the performance with our proposed framework. The empirical outcomes highlight the critical role of the planning module (TCPM) in learning an efficient search policy in partially observable environments. Please refer to lines 441- 460 in the manuscript for a discussion of this.

Q2: My first question is about the practicality in the real world, if the author can convince me that this work can be applied in the real world, then I won't have any further questions. Otherwise, I think that the current technical framework itself lacks sufficient novelty. I know we can find detailed differences between papers, but hope the author could clarify the core technological novelty of this paper.

A2: Real-world practicality: It is not our view that novel research has to be able to run in a fully online, real-world setting -- that gap is quite unrealistic to be expected to be filled within a single paper that starts off at a new idea and establishes important foundations, in our opinion. Our method is at this research stage not yet implemented to run e.g. on a physical drone, and there are naturally quite a few steps to transition the code to run appropriately in such a real-world setup. Challenges to tackle include e.g. (i) transfer models trained in emulated satellite-data setting to use drone imagery; (ii) handle imperfect drone actuation and blurry imagery due to movements; (iii) ability to run on-device using limited compute. As for (iii), for example, one can leverage recent rapid advancements in diffusion models that make them more compute and time-efficient by reducing the number of reverse diffusion steps (such as [1,2] ). In principle, however, the approach could be proof-of-concepted even on actual drones, especially if one at the early stage allows for non-real-time processing.

Novelty of our work: We would like to highlight the main contributions and novelties of our work:

(1) We introduce the novel task of target-conditioned visual active search in partially observable environments. This is different to prior work which assumes full scene observability.

(2) To address this task, we propose a novel framework that combines a diffusion-based Conditional Generative Module with an RL-based target-conditioned Planning Module to reason effectively in these partially observable environments. We design a novel reward function to guide the TCPM module in mastering an efficient search strategy that requires balancing exploration (query regions to gather important information about the environment) and exploitation (query regions based on the current belief to maximize target discovery) in partially observed scenes.

(3) Additionally, our proposed inference strategy enables the DiffVAS framework to handle search tasks involving multiple target categories, which is different from prior work that focuses on single-target settings.

Our primary focus in this work has been to lay the foundation for the novel and real-world relevant task of Target-Conditioned Visual Active Search in Partially Observable Environments, which we term TC-POVAS. We hope our work serves as a catalyst for future advancements in the field of "Active Search in Partially Observable Environments".

References:

1. Frans, Kevin, et al. "One Step Diffusion via Shortcut Models." arXiv preprint arXiv:2410.12557 (2024).

2. Delbracio, Mauricio, and Peyman Milanfar. "Inversion by direct iteration: An alternative to denoising diffusion for image restoration." arXiv preprint arXiv:2303.11435 (2023).

Thanks for your detailed reviews and thoughtful questions. Below, we address all your comments.

Q1 Problem setting to be optimized.

A1 There are certainly variations of our task setup that could be explored, each with its own trade-offs and assumptions (for example, a third setting would be one where the agent can only move to an adjacent position in each action). The task setup you suggest, where the agent gets to observe all patches between its previous and current patch, has the risk of leading to fairly uninteresting search behaviors, where a (close-to) optimal strategy may be to always move to a maximally distant uncovered location (since on average, that implies as many revealed subareas as possible, which in turn leads to on average more targets being discovered). Furthermore, when the agent makes a movement that is not vertical or horizontal, there are multiple possible paths that could be filled in, thus becoming a bit ill-defined how to do the fill-in in many cases. We therefore believe our current task formulation is reasonable, but we do recognize that parts of the suggestion could be incorporated in future work, although arguably the "intermediate" revealed sub-areas ought to be blurred due to movement.

Q2 Details on training and testing are missing.

A2 We discuss the implementation details in Appendix section A4. Also, in the main paper starting on Line 383, we write: "Both xView and DOTA are satellite image datasets, with roughly 3000 px per dimension and representing approximately 60 object categories. We use 50%, 17%, and 33% of the large satellite images to train, validate, and test the methods, respectively". For convenience, we provide the anonymous link to the script for training and inference of DiffVAS in Appendix section A4 of the updated draft. We are committed to open-sourcing the GitHub link upon acceptance, including all model weights, to facilitate an easy reproduction of all our experimental results. We would be happy to provide any further details about our experiments if you require any additional specifics or clarification.

Q3 Details about inference with E2EVAS and MPS-VAS.

A3 During inference, at each search step, we provide only the partially observed glimpses to the policy network, rather than the full search image. Specifically, we input an image with black patches representing unobserved areas and the ground truth content for the already observed regions. Based on the policy's action, we then update the input image by adding the newly revealed patch, which corresponds to the patch selected by the policy for querying. Aside from that, the E2EVAS and MPSVAS remain unchanged from the original version.

Q4 Motivate the experimental setup in multi-target and zero-shot setting.

A4 In both xView and DOTA, due to the large combinatorial space of possible category subsets, we selected only a representative subset for evaluation. Moreover, certain classes are extremely rare within the dataset. For instance, images containing at least one instance of class Cement-Mixer appear only once. Consequently, including such classes as targets is not meaningful, as evaluating performance based on a single image does not provide robust analysis/results. Furthermore, regarding the multi-target results, in addition to those reported in Table 2 of the main paper, Table 9 in the Appendix includes 3 more multi-target combinations, bringing the total to 6 combinations in this work.

The zero-shot results are presented starting around line 494 in the main paper. In this section, we train models on DOTA using categories that do not appear in xView. Then, in Table 7, we report the zero-shot performance on the unseen xView classes. These additional details have been included in Appendix A9 of the updated draft.

Q5 How do you sample episodes during training?

A5 The details of our training and evaluation settings are provided in Appendix A4. During training, we randomly sample an image from the training set and select a random budget within the range {1 to N-1} (where N is the number of grids). We then extract the set of target categories if at least one instance of each category is present in the image, based on the ground truth annotations (which also include the precise locations of these target objects as a bounding box). Once the target category set is determined, we randomly select a category from this set to train our policy. The rationale for using a random-budget and random-target is to ensure that the policy learns to be robust against the target category and budget constraint. Across various experiments, we demonstrate that DiffVAS consistently outperforms the baseline across different budgets and target categories. These additional details have already been included in Appendix A10 of the updated draft. Moreover, we would be happy to incorporate any other specific training details you suggest for the final version.

Dear Reviewer yyn7,

As the discussion period approaches its conclusion, we would like to follow up to ensure that all your concerns have been fully addressed. If you require any further clarifications, please do not hesitate to let us know, and we will be glad to provide them promptly. Thank you for all your time and effort.

Regards,

Authors

Dear Reviewers,

Since we are now past the halfway point of the discussion period, we want to ensure that our responses have fully addressed your concerns. Please let us know if any further clarification or details are needed, and we would be happy to provide them as quickly as possible. If you believe our response has adequately addressed the concerns you raised, we kindly ask you to consider the possibility of raising the score. Thank you for your time.

Regards,

Authors