acceptodds
Under review as a conference paper at ICLR 2027

DCA-MoE: Spatially Adaptive Cross-Layer Fusion and Density-Routed Experts for Crowd Counting

Abstract

Crowd counting must recover reliable local density under severe variations in perspective, head scale, occlusion, and background clutter. Although modern counting objectives provide strong spatial supervision, many multi-level decoders still use spatially invariant feature fusion and apply one receptive-field pattern to every location. We propose DCA-MoE, a framework that makes both decisions content dependent while retaining a frozen DINOv3 encoder. Spatially Adaptive Layer Fusion (SALF) predicts position-wise weights over four aligned backbone features, and Density-Routed Multi-Receptive-Field Experts (DR-MoE) assigns each location a soft mixture of local, mid-range, and large-context residual experts. An EBC-style head reconstructs block density, while DMCount supervision and an auxiliary routing-balance term train the decoder without updating the backbone. On the NWPU-Crowd validation split, the strongest paired configuration, based on DINOv3 ViT-L/16, obtains 31.7 MAE and 72.2 RMSE; the matched ViT-B/16 full model obtains a paired 32.2/75.9. Cross-dataset results remain mixed, and several component baselines currently report independently selected minima from a single seed. The evidence therefore supports the feasibility of spatially adaptive fusion and routing, while broader paired and multi-seed evaluation remains necessary for causal attribution.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
rowd counting, density estimation, spatially adaptive fusion, mixture of experts, DINOv3, blockwise classification

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_axc7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DCA-MoE, a decode that retains multi-layer features from frozen DINOv3 encoder, aggregate them through spatial fusion and across multiple receptive fields. The method is evaluated on NWPU-Crowd and shows improvements over several previous methods.

Critical strengths
  • The method demonstrates quantitative improvements on NWPU-Crowd compared to several existing methods.
Critical weaknesses
  • The technical novelty is limited. The proposed components are primarily learnable softmax weights on concatenated features and multi-branch feature processing. The paper does not clearly convey what these designs offer beyond existing feature fusion approaches.
  • Though the paper claims a motivation of using frozen DINOv3 features, the experiments fail to validate it, as the only model using same backbone size is CLIP-EBC. However, CLIP-EBC utilizes CLIP as backbone instead of DINOv3. So there’s no direct comparison.
  • The presentation is incomplete and lacks clarity. For example, Fig. 1 only provides text to explain each block.
Questions for authors

Please see Weaknesses

Reviewer_VV5rOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DCA-MoE, a crowd counting framework combining Spatially Adaptive Layer Fusion (SALF) and Density-Routed Multi-Receptive-Field Experts (DR-MoE) with a frozen DINOv3 backbone. SALF adaptively fuses multi-level features, while DR-MoE aggregates experts with different receptive fields through spatial routing. An EBC-based head is used for density estimation and counting. Experiments on the NWPU-Crowd validation set show improved performance over several existing methods, with component ablations and qualitative results provided for validation.

Critical strengths
  1. The paper addresses the practical challenges of scale variation and uneven crowd density. The proposed framework provides a reasonable combination of adaptive feature fusion and multi-receptive-field modeling to handle spatial variations.
  2. The method achieves competitive performance on the NWPU-Crowd validation set, and the combined SALF and DR-MoE configuration demonstrates improvements over the fixed-fusion baseline, suggesting potential benefits from their integration.
Critical weaknesses
  1. The novelty of the paper is limited at both the problem and methodological levels. Spatially adaptive feature fusion and multi-receptive-field modeling have been extensively explored in crowd counting and related vision tasks. The proposed SALF and DR-MoE modules mainly combine existing attention-based fusion and mixture-of-experts mechanisms, without introducing a clearly differentiated learning principle or demonstrating a substantial methodological advance.
  2. The experimental evaluation is insufficient to substantiate the claimed contributions. Results are reported only on the NWPU-Crowd validation set, and the ablation study relies on a single random seed with inconsistently selected checkpoints. Moreover, the individual modules yield only marginal improvements, while their larger combined gain lacks further explanation. The absence of routing analyses also makes it difficult to verify the claimed density-dependent expert specialization.
Questions for authors
  1. What fundamentally distinguishes the proposed SALF and DR-MoE from existing spatially adaptive fusion, dynamic convolution, and multi-branch expert architectures? Beyond combining these established mechanisms, what is the main methodological contribution?
  2. Can the authors provide evidence that DR-MoE learns density-dependent expert specialization rather than simply performing generic spatial attention-based feature fusion?
  3. Why do SALF and DR-MoE individually provide negligible improvements, while their combination produces a much larger gain? Can this result be verified through fully paired, multi-seed experiments?
  4. Can the authors provide evaluations on additional crowd counting datasets to establish the generalizability of the method?
Reviewer_YpogOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes DCA-MoE, a decoder for crowd counting that adaptively selects feature depth and receptive fields at each spatial position while keeping the DINOv3 encoder frozen. The method combines SALF for cross-layer feature weighting, DR-MoE for residual expert fusion, and an EBC-style objective for density and count estimation. Experiments on the NWPU dataset show some promising results, but the evaluation is limited and the novelty and presentation of the proposed method are not sufficiently clear.

Critical strengths

The method attempts to adapt feature depth and receptive fields according to different spatial locations, which may be useful for handling scale and density variations in crowd counting.

Critical weaknesses
  1. The experimental evaluation is conducted on only one dataset, which is insufficient to demonstrate the generalization capability of the proposed method.

  2. The performance on the NWPU test set is not reported, making it difficult to fairly compare the method with existing state-of-the-art approaches.

  3. The contribution of the proposed method is not clearly explained, and the overall writing and presentation of the method should be substantially improved.

  4. The novelty of the proposed method is limited, as the main components are largely based on existing feature fusion, mixture-of-experts, and density estimation techniques.

Questions for authors
  1. Why is the method evaluated only on the NWPU dataset? Can the authors clarify whether the method is expected to generalize to other crowd counting datasets?

  2. Why is the performance on the official NWPU test set not reported?

  3. What are the key technical differences between SALF and existing adaptive cross-layer feature fusion methods?

  4. What are the main differences between DR-MoE and existing mixture-of-experts or dynamic feature fusion approaches?

  5. Can the authors more clearly explain the main methodological contribution and why the proposed combination should be considered novel?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.