acceptodds
Under review as a conference paper at ICLR 2027

CPCNet: Cross-Depth Pyramid Construction for Underwater Object Detection

Abstract

Adapting pretrained vision transformers to underwater object detection requires constructing a multiscale pyramid from encoder features that share a common spatial grid. How features from different encoder depths are combined during this construction determines the representations supplied to the detector. We propose CPCNet, an underwater object detector built around a Cross-depth Pyramid Constructor (CPC). CPC jointly aggregates four projected DINOv3 states on their native grid to form a shared central representation. This center feeds a finer branch and a progressively coarser chain, with depth-specific bypasses providing direct encoder inputs to the branches. In single-run evaluations, CPCNet achieves 69.66, 65.27, and 52.18 AP on DUO, RUOD, and UTDAC, respectively. Across all three underwater datasets, CPC achieves higher AP than approximately parameter-matched independent routing and higher AP with fewer operations than source-complete parallel construction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Underwater Object Detection, DINOv3, Vision Transformers, Cross-Depth Feature Aggregation, Feature Pyramid Construction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rhESOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CPCNet for underwater object detection. Its main component, the Cross-depth Pyramid Constructor (CPC), aggregates four same-resolution DINOv3 encoder features into a shared central representation. This representation is propagated to finer and coarser stages, with depth-specific bypasses supplying additional encoder features. The resulting pyramid feeds an FPN–PAN neck and a TOOD-style detection head. Evaluation covers DUO, RUOD, and UTDAC, with additional COCO results. The paper compares alternative construction topologies and provides component ablations, feature probes, qualitative examples, and degradation analyses.

Critical strengths
  1. Useful comparisons within a common detection framework. The paper goes beyond comparing complete detectors by keeping the DINOv3 backbone, downstream neck, head, and training recipe fixed across several construction variants. The parameter-matched direct-routing and source-complete parallel baselines are particularly relevant for assessing whether the proposed feature organization adds value.

  2. Consistent results across three underwater datasets and transparent analysis. CPC achieves higher AP than both principal construction controls on DUO, RUOD, and UTDAC. The paper also explicitly acknowledges limitations, including selected qualitative examples, nonuniform probe improvements, and confounded ablations. This transparency helps distinguish measured findings from proposed explanations.

Critical weaknesses
  1. Limited methodological novelty. CPC mainly combines established concatenation, convolutional mixing, resampling, and shortcut operations. The paper explains its structural differences from existing approaches, but does not sufficiently establish a substantial advance over related cross-layer aggregation and ViT adaptation methods.

  2. Small gains over the closest baseline without statistical validation. CPC improves over source-complete parallel construction by only 0.32, 0.39, and 0.44 AP on DUO, RUOD, and UTDAC, respectively. All results come from single training runs. Without repeated runs or uncertainty estimates, it is unclear whether these margins reflect a reliable advantage.

  3. The ablations do not isolate the proposed mechanisms. Removing recursive propagation also changes encoder-source coverage, while removing bypasses changes mixer input widths, parameter counts, and shared-operator calls. These experiments compare complete configurations but cannot establish whether propagation or bypass connections themselves explain the improvement.

  4. Important controlled baselines are missing. The final-state simple pyramid has substantially fewer parameters and operations than CPC. A parameter-matched final-state baseline is needed to separate the benefit of cross-depth information from additional processing capacity. Closely related ViT feature-adaptation methods are also discussed mainly through architectural descriptions rather than direct comparisons under the same framework.

  5. The computational cost is substantial, but actual detection speed is unreported. On DUO, CPC increases computation from 322.07 to 409.66 GFLOPs relative to the simple pyramid, while improving AP by 1.72 points. DINOv3 feature extraction may introduce additional latency compared with conventional detection backbones, but the paper reports neither FPS nor measured inference latency. The practical accuracy–speed trade-off therefore remains unclear.

  6. The evidence connecting CPC to underwater-specific challenges is weak. The motivation emphasizes background similarity, low contrast, and overlapping targets, but the proposed structure is a general feature-fusion design. The degradation analysis evaluates the full CPCNet without corresponding baseline comparisons, so it does not establish that CPC specifically improves robustness to these conditions.

Questions for authors
  1. What methodological insight distinguishes CPC from existing cross-layer aggregation and ViT adaptation approaches beyond a different arrangement of standard operations?

  2. Can you isolate stage propagation while preserving encoder-source coverage, and isolate bypass connections while controlling capacity and operator sharing? The current ablations change several factors simultaneously.

  3. What are the batch-one inference latency and FPS under identical hardware and precision settings? Please compare CPCNet with the simple-pyramid and source-complete parallel baselines, and separate backbone time from the remaining detector computation.

  4. Does CPC outperform the construction baselines under the same blur and contrast degradations? Are its gains larger on low-contrast or overlapping targets? Such comparisons would clarify whether the proposed design addresses the underwater challenges motivating the paper.

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

In this paper, the authors proposed CPCNet, which adapts same-grid DINOv3 features to underwater object detection according to joint cross-depth pyramid construction. It combines shared cross-depth aggregation, stage propagation, and direct depth-specific inputs to construct a multiscale detection pyramid from same-grid encoder state. Different experiments on three datasets demonstrate its performance.

Critical strengths

-This work is reasonably well-motivated and proposes a targeted approach to Transformer feature organization and multi-scale representation learning. - The ablation studies cover computational overhead, feature responses, and failure case analysis, demonstrating a reasonable degree of experimental thoroughness and research rigor. - The figure is clearly presented, effectively illustrating the proposed framework and its key differences from the conventional FPN-PAN architecture, facilitating readers' understanding of the design. - The paper is relatively well-structured and clearly presented, making it easy for readers to follow and understand.

Critical weaknesses

- Limited novelty. The methodological novelty appears somewhat limited relative to the conference's expectations. Although the proposed CPC module differs from existing methods in its architectural design and feature connections, its underlying mechanism remains largely similar, offering limited methodological innovation.

- The paper mainly distinguishes CPC from existing methods through architectural descriptions and empirical comparisons, while providing limited theoretical justification for its design choices and insufficient analysis of its underlying mechanisms.

Questions for authors

- Limited performance gains relative to computational cost. As shown in Table 2, CPC increases parameters by 21.0% (52.56M to 63.59M) and computational cost by 27.2% (322.07 to 409.66 GFLOPs) over Final-state simple pyramid, while improving AP by only 1.72%. Compared with Source-complete parallel, CPC achieves merely a 0.32% AP gain with similar parameters. These results raise concerns about its computational efficiency and practical benefits.

- On the DUO and RUOD datasets, the proposed method yields only marginal improvements over existing detectors, raising concerns about its performance advantages and overall effectiveness.

-Additional ablation studies on key components are recommended, such as selectively disabling bypass connections at different scales, to better assess their individual contributions and justify the design choices.

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

This paper studies underwater object detection task. The paper proposes a Cross-depth Pyramid Constructor (CPC) to aggregate feature from Dino v3 to improve feature fusion for underwater object detection. CPC consists of a shared cross-depth feature aggregation module, recursive multi-scale feature construction, and depth-specific bypass connections. Experimental results on three underwater object detection datasets, DUO, RUOD, and UTDAC, demonstrate the effectiveness of the proposed method.

Critical strengths

The model achieves improved underwater object detection performance through architectural refinements. Although the proposed method's novelty is limited, the performance improvements may still offer practical benefits for real-world engineering applications.

Critical weaknesses

Feature pyramids are a well-established architectural paradigm in object detection. Although the proposed CPC introduces cross-depth feature aggregation, recursive stage propagation, and depth-specific bypass connections, these components appear to extend or combine existing feature fusion and multi-scale feature construction techniques rather than introduce fundamentally novel architectural principles. The paper does not sufficiently clarify how CPC fundamentally differs from existing multi-level feature fusion methods, nor does it adequately justify the necessity of its specific design beyond the observed empirical improvements. Consequently, the paper does not sufficiently establish the novelty of the proposed architecture.

Questions for authors
  1. The paper states that the encoder, CPC, neck, and detection head are jointly trained. How are these weights determined?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.