acceptodds
Withdrawn from ICLR 2027

Open-Text Underwater Detection: Grounding Rich Descriptions Across Objects and Context

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Underwater object detection becomes challenging when object appearance alone is insufficient to distinguish target instances, while relations and environmental context provide complementary evidence. We study open-text underwater detection, where rich descriptions of appearance, relations, and context are used to localize all matching instances. We introduce AquaGround, to our knowledge the largest benchmark for open-text underwater detection, containing 224K images, 893K object instances, 2,363 categories, and 7.52M query records. AquaGround provides four progressively enriched query levels and five structured environmental attributes, with each query explicitly mapped to all valid matching instances. We further propose AquaDetect, a cue-guided open-text underwater detector that combines target semantics, underwater context, and detail–redundancy cues to guide progressive visual aggregation while preserving spatial correspondence for localization. Experiments on AquaGround show that AquaDetect reaches 58.94 AP, outperforming the SOTA method by 5.22 points.

Primary area:
datasets and benchmarks
Keywords:
Open-Text Detection, Underwater Vision, Visual Grounding, Vision-Language Models, Open-Vocabulary Object Detection, Benchmark Dataset

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

This paper studies open-text underwater detection, where a natural-language description may refer to multiple objects through appearance, relations, and environmental context. It introduces AquaGround, a benchmark containing approximately 224K images, 894K instances, 2,363 categories, and 7.52M image–query records. The annotations provide four progressively enriched query levels and structured environmental attributes. The proposed AquaDetect uses six guidance scores to select aggregation centers, progressively merge visual tokens, and recover spatially organized features for detection. Experiments report 58.94 AP, exceeding OTA-Det-M by 5.22 points, together with improvements in visual grounding and component ablations.

Critical strengths
  1. AquaGround is a potentially valuable benchmark for open-text underwater detection. It combines object annotations, multi-level descriptions, environmental attributes, and one-to-many query–instance associations at a substantial scale.

  2. The proposed semantic-guided token aggregation and spatial recovery form a coherent design for language-conditioned detection. The reported improvements over OTA-Det-M and the component ablations provide encouraging evidence of its effectiveness.

Critical weaknesses
  1. The inference setting is insufficiently clear. Appendix C.1 states that AquaDetect uses dataset-provided auxiliary descriptions and structured conditions during inference. It is unclear how these object-linked annotations are selected without knowing the target instances, and whether all baselines receive equivalent information. This makes it difficult to attribute the gains to the proposed method.

  2. The reliability of the final query–instance associations needs stronger validation. Descriptions and matching sets are largely constructed automatically, but the reported expert checks do not separately quantify incorrect matches and missing valid targets. These errors could directly affect both training and evaluation.

  3. The independence of the data splits is not established. The duplicate-removal procedure targets near-identical images but does not exclude related scenes or adjacent video frames across splits. Consequently, generalization to independent underwater scenes remains uncertain.

  4. Some figures, particularly Figure 2, have an AI-generated appearance and differ noticeably from the other figures in color palette, line thickness, and typography. Although their origin cannot be determined from appearance alone, the inconsistent styling reduces visual coherence. Please improve their clarity and standardize the presentation across all figures.

Questions for authors
  1. Please clarify the exact inference inputs using a concrete example. How are object-linked auxiliary descriptions and environmental conditions selected without using ground-truth boxes or query–instance associations?

  2. Do all baselines receive the same auxiliary information, including structured conditions? Please provide an image-plus-query comparison between AquaDetect and the strongest baseline to separate architectural gains from the benefit of additional inputs.

  3. Has the final query–instance mapping been independently checked by humans? What are the rates of incorrect matches and missing valid targets, particularly for relational and multi-instance queries?

  4. Are images from the same video, dive, or acquisition sequence kept within one split? If sequence information is available, can the authors report performance on a sequence-disjoint subset?

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

This paper introduces AquaGround, which is claimed to be the largest benchmark for open-text underwater object detection to date. Furthermore, the authors propose AquaDetect, a detection framework that effectively improves underwater object detection performance. Overall, this work offers valuable contributions to underwater object detection research.

Critical strengths

- This work is well-motivated and introduces AquaGround, a dataset featuring diverse object instances and categories, along with rich semantic associations across categories, appearances, relationships, environments, and spatial locations, making a valuable contribution to underwater object detection research. - Tailored to the characteristics of AquaGround, the authors develop AquaDetect, a detection framework that leverages multi-level cues to guide information fusion and improve detection performance. -Extensive experiments on open-text detection and visual grounding tasks demonstrate the value of the AquaGround dataset and the effectiveness of the proposed AquaDetect framework. - The paper is relatively well-structured and clearly presented, making it easy for readers to follow and understand.

Critical weaknesses

- Although the AquaGround dataset incorporates rich and detailed descriptions of object appearances, relationships, and contextual information, the necessity and effectiveness of these descriptions for open-text detection and visual grounding tasks have not been sufficiently demonstrated.

- Although AquaDetect incorporates multiple cues, including Category, Object Attributes, Relations & Environment, and Background & Position, the specific roles of these cues in guiding information fusion remain insufficiently explained, and their effectiveness has not been adequately validated through experiments.

Questions for authors

- It is recommended to include visualizations of both successful and failure cases to further illustrate the working mechanisms of key modules, such as the Cue-Guided Router, in AquaDetect and their impact on detection performance.

- Some statistical results appear to be inaccurate. For example, in Table 10 of the Appendix, the Big category contains 271,385 instances out of a total of 893,799, corresponding to 30.36%. The authors are encouraged to carefully verify the reported statistics to ensure their accuracy and consistency.

- Fairness of detector comparisons. Several compared detectors, such as BSR5-DETR, YOLO-World, and OTA-Det-M, do not explicitly utilize additional semantic cues, including Relations & Environment and Background & Position, whereas AquaDetect is specifically designed to exploit such information. Thus, its performance gains may partly result from these additional cues rather than the framework itself. Moreover, these cues may be difficult to obtain accurately during real-world inference, raising concerns about comparison fairness and practical applicability. The authors should clarify how these cues are obtained during inference and conduct comparisons under consistent input conditions.

- Some formatting and writing inconsistencies remain. For example, capitalization is inconsistent in headings such as “Open-text detection Results” and “Top-1 visual grounding Results”. Additionally, step titles such as “Refine Expressions Within each image” lack ending periods. The authors are encouraged to carefully review the manuscript to ensure consistent formatting and writing conventions.

Reviewer_BTtiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies open-text underwater detection: locating all instances that satisfy a rich description. AquaGround contains 224,632 images, 893,799 objects, 2,363 categories, and 7,522,059 image-query records. Queries span category, attributes, relations/environment, and background/position. 1,490,007 queries refer to multiple targets. AquaDetect combines DINOv3 TinyPlus with frozen SigLIP 2. Six guidance signals covering target semantics, context, and detail/redundancy drive three stages of visual-token aggregation. Spatial correspondence, TRA, and a CNN branch then recover multi-scale features. On validation data, it achieves 58.94 AP, 5.22 points above OTA-Det-M, and 37.47 [email protected], 2.89 points above that column's strongest baseline, Grounding DINO. These results use dataset-provided auxiliary annotations at inference, a material condition for interpreting the comparison.

Critical strengths
  1. Useful supervision beyond scale. AquaGround links rich object descriptions and environmental attributes to stable object identities and explicitly maintains the set of targets for each query. This structure supports systematic study of how language distinguishes multiple objects in complex underwater scenes.
  2. The internal ablations support the design under the reported input conditions. Removing context, detail/redundancy, or TRA reduces AP by 2.17, 2.91, and 2.16 points. Direct embedding with the same auxiliary information yields 55.06 AP versus 58.94 for the full model. [email protected] also improves from the strongest baseline's 34.58 to 37.47, supporting better strict localization.
Critical weaknesses
  1. Inference inputs and attribution remain unresolved. C.1 supplies auxiliary descriptions and environmental conditions from annotations at inference; absolute-position fields are precomputed from object boxes. These provide object-level semantic/location hints beyond the current query, although GT coordinates are not directly fed to the detector. Equivalent access across baselines is not established by the statement “identical textual inputs”. The same-input direct-embedding control is useful, but does not establish an advantage with image and query alone. Input-matched comparisons are essential to assess practical applicability.
  2. Final target-set validity and evaluation robustness lack sufficient evidence. The approximately 1% expert check addresses generated semantics; final query-instance precision, recall, and omissions are not independently quantified (B.1/B.4/B.7). It is unclear whether matching constraints filter attributes absent from a query. Protocol A retains only 100 predictions across all queries of an image before scoring image-query categories, coupling AP to query count and cross-query score calibration (D.1). Per-query budgets, stratified results, and an independent final-set audit are needed. Main results use validation data; held-out test results, Protocol B’s evaluated population, and checkpoint-selection rules also need clarification.
Questions for authors
  1. Clarification and decisive check: For every method, list the training/inference input fields and how auxiliary descriptions and conditions are associated with queries/objects. How do results change with only the image and current query, or automatically generated auxiliary information? Specify Eq. (8)'s valid mask and the query bundle used at inference (C.1).
  2. Controlled comparison: Explain direct embedding and its budget. With backbone, detector, training, and auxiliary inputs matched, compare no routing, simple similarity-based merging at the full model's token budget, and the full method (Table 3; Appendix C).
  3. Annotation check: State the final matching rules, including constraints on fields absent from queries. Can an independent audit quantify incorrect and missing links, sample size, and sampling unit by query level and single/multiple targets (B.1/B.4)?
  4. Evaluation check: Recompute the main table with independent per-query maxDets=100 and stratify by query count, level, and target multiplicity. Specify Protocol A/B splits, sample counts, Category inclusion, and checkpoint-selection rules; report the fixed model on the test set (Secs. 5.1-5.2; D.1).

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

Related papers

Discussion (0)

Sign in to comment.