acceptodds
Withdrawn from ICLR 2027

ReasonRSOS: Reasoning Salient Object Segmentation in Remote Sensing Images

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Remote sensing salient object segmentation (RSI-SOS) aims to segment the most visually salient objects in remote sensing images, and has achieved remarkable progress on conventional benchmarks. However, existing methods remain largely visual-driven and therefore lack the ability to interpret implicit user intent in complex geospatial scenes. To extend the research boundary, we introduce a new task, termed Reasoning Salient Object Segmentation in Remote Sensing Images (R-RSISOS). This task requires models to segment salient remote sensing objects from implicit text queries by leveraging geospatial context, complex reasoning, and world knowledge. To support this task, we construct ReasonRSOS, a benchmark containing 8,265 image-query-mask triplets. We further introduce IRSS, a challenging subset containing multi-salient-target scenes for high-level remote sensing reasoning.We also propose RISA, a Remote Instructed Salient Assistant that segments salient remote sensing objects according to implicit text queries. Extensive experiments show that RISA achieves superior performance on ReasonRSOS and demonstrates promising generalization ability on related tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Language, Multi-modal, Salient Object Segmentation, Remote Sensing Images

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fQRaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces reasoning salient-object segmentation in remote sensing and the ReasonRSOS benchmark of 8,265 image-query-mask triplets, including the new IRSS subset. RISA combines visual encoding, query-aware context selection, an MLLM, and mask decoding. Experiments report improvements over multimodal segmentation baselines and evaluations on related segmentation tasks.

Critical strengths
  1. The benchmark targets a useful gap between visual saliency and intent-dependent target selection in overhead imagery. The new IRSS scenes could be valuable for studying ambiguity among multiple plausible targets.

  2. The annotation protocol includes three-expert verification, and Appendix G adds blinded human evaluation of explanations on 500 test samples. Table 2 also reports consistent, although modest, gains over a strong 7B remote-sensing baseline.

Critical weaknesses
  1. The benchmark does not yet isolate dependence on the query. Existing saliency images retain their original masks and receive one implicit query each (Sec. 3), leaving a plausible image-only shortcut. Although Appendix E discusses conventional saliency baselines, Table 2 does not quantify such a control. Matched image-only/query-shuffling controls and controlled cases with different targets for the same image are important to establish that the benchmark measures the intended reasoning capability.

  2. The central method is insufficiently specified and does not map clearly to the supplied code. On p. 6, Xc is a top-K subset of text tokens Xq, and the MLLM receives [Xc; Xq]; it is unclear where visual value features enter this representation. The supplied training instructions and Phi-based implementation also differ from the LLaVA-1.5 7B/13B setup in Sec. 5.1, and the documented training entry is absent. These gaps prevent verification of what produced Table 2 and of the contribution beyond established SEG-token decoding.

Questions for authors
  1. Is Xq in the QACS formula already visually fused? Please give the exact tensors and implementation entry corresponding to the context passed to the MLLM.

  2. Can you provide the exact 7B/13B configurations and matched image-only or shuffled-query results, especially on IRSS?

  3. Were the Table 2 baselines trained with the same in-domain and auxiliary data? Which additional fine-tuning, if any, was used for each Table 5 benchmark?

Reviewer_h9gkOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses salient object segmentation in remote sensing images (RSI-SOS). It introduces R-RSISOS, a new task that requires segmenting salient remote sensing objects from implicit text queries by leveraging geospatial context, complex reasoning, and world knowledge. A new dataset ReasonRSOS is constructed; it contains 8,265 image–query–mask triplets, along with IRSS, a challenging subset containing multi-salient-target scenes for high-level remote sensing reasoning. It further proposes RISA, a Remote Instructed Salient Assistant that segments salient objects according to implicit text queries. Extensive experiments show that RISA achieves superior performance on ReasonRSOS and demonstrates promising generalization ability on related tasks.

Critical strengths
  1. This paper formalizes it as a new task, R-RSISOS, which requires combining geospatial context, complex reasoning, and world knowledge to segment salient targets from implicit text queries.

  2. This paper constructs ReasonRSOS, a benchmark of 8,265 image–query–mask triplets with reasoning annotations, and further introduces IRSS, a challenging subset specifically targeting multi-salient-target scenes where visual saliency alone cannot determine the intended target.

  3. It proposes RISA, a reasoning-based segmentation framework that identifies intention-consistent salient targets, and extensive experiments show that RISA achieves superior performance on ReasonRSOS with promising generalization ability on related tasks.

Critical weaknesses
  1. How sensitive are the results to the scale of the training set — would increasing or reducing the number of training images substantially change the reported performance?
  2. The gap between the 13B and 7B results is not large. What is the main bottleneck of this task — language reasoning capacity, visual grounding, or the scale of annotated data?
Questions for authors

Please see the weaknesses.

Reviewer_qNK7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ReasonRSOS, a benchmark of 8,265 image-query-mask triplets for implicit reasoning-based salient object segmentation in remote sensing imagery. It also proposes RISA, a vision-language segmentation framework, and evaluates its performance against existing methods on the proposed benchmark and related segmentation tasks.

Critical strengths

Extending remote sensing segmentation from visual saliency toward implicit user-intent understanding is practically relevant, particularly for complex scenes containing multiple candidate targets.

Critical weaknesses
  1. Prior works, including LISAt (NeurIPS 2025), SegEarth-R1 (2025), and SegEarth-R2 (CVPR 2026), have already investigated implicit-query reasoning segmentation in remote sensing imagery. The manuscript does not sufficiently establish the fundamental differences or additional research value of reasoning-based salient object segmentation relative to these existing formulations.

  2. Most benchmark samples are derived from existing salient object segmentation datasets by adding implicit queries to predefined masks. Without appropriate query-dependent controls, it is difficult to determine whether the benchmark genuinely measures reasoning ability rather than conventional saliency recognition.

Questions for authors
  1. Could the authors clarify the fundamental differences between ReasonRSOS and existing geospatial reasoning segmentation benchmarks, particularly GRES (LISAt) and EarthReason (SegEarth-R1), and explain what new reasoning challenges are introduced by the proposed salient-object formulation?

  2. Could the authors provide evidence that the segmentation predictions genuinely depend on implicit queries, for example through image-only baselines or query replacement experiments in multi-target scenes? This would help establish the necessity of reasoning for the proposed task.

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

Related papers

Discussion (0)

Sign in to comment.