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