acceptodds
Withdrawn from ICLR 2027

MFCA-Net: Multi-Frequency Content-Adaptive Network for UAV Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Object detection performance can be improved through complementary techniques such as data augmentation, loss optimization, and image feature enhancement. Inspired by conventional image filtering, we propose a Multi-Frequency Content-Adaptive Network (MFCA-Net), which consists of Fixed-Frequency Prior Extraction (FFPE), Multi-Frequency Prior Representation (MFPR), and Content-Adaptive Frequency Fusion (CAFF). MFPR extracts three prior features using Gaussian filtering, Difference of Gaussians (DoG), and an eight-neighbor Laplacian filter. CAFF generates frequency-specific weights from the RGB and prior features, fuses them into a single-channel representation, and concatenates it with RGB as the input to the detector. We evaluate MFCA-Net with YOLOv8 and MaxViT backbones on VisDrone-DET and DroneVehicle. MFCA-Net introduces only 11,751 additional parameters while improving detection performance across different backbones and datasets. In particular, YOLOv8n+MFCA-Net improves mAP by 1.2% and mAPs by 0.9% on VisDrone-DET. These results demonstrate that the proposed frequency-based feature enhancement can improve object detection with a small increase in model complexity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fixed Filters, Prior Features, Content-Adaptive Fusion, Multi-Frequency Representation, Object Detection, YOLOv8

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

This paper proposes MFCA-Net for UAV object detection. It extracts low-, mid-, and high-frequency priors from the input using Gaussian filtering, Difference of Gaussians, and Laplacian filtering, respectively. A lightweight Content-Adaptive Frequency Fusion module predicts spatial weights for these frequency components and fuses them into one additional channel, which is concatenated with RGB and fed into standard detectors. The method is evaluated with YOLOv8n and MaxViT-T on VisDrone-DET and DroneVehicle.

Critical strengths

1.The method is simple and easy to integrate into existing object detectors. 2.The paper provides ablations on different fixed filters, frequency combinations, and detection scales.

Critical weaknesses

1.Technical novelty is quite limited. MFCA-Net mainly combines standard Gaussian, DoG, and Laplacian filters with a small learned module for spatially adaptive fusion. This is closer to a straightforward integration of handcrafted image priors and learned weighting than a substantial methodological advance.

2.With a stronger backbone, the improvement becomes marginal. On VisDrone-DET, MaxViT-T improves only from 24.6 to 24.7 mAP, while on DroneVehicle the gain is from 53.0 to 53.3. This makes it unclear whether the proposed frequency priors are generally useful or mainly compensate for the limited representation capacity of the smaller YOLOv8n backbone.

3.Efficiency is also somewhat overstated. Although the method adds only 11,751 parameters, YOLOv8n GFLOPs increase from about 4.06 to 8.67, and FPS drops from about 85.7 to 72.7. The method is lightweight in terms of parameter count, but not in overall computation.

4.It is not well isolated where the gain actually comes from. The paper does not compare against similarly lightweight learned alternatives, such as a small convolutional preprocessing module or an attention module without handcrafted frequency decomposition. As a result, the current experiments do not clearly show that the improvement is specifically due to the proposed frequency priors.

Questions for authors

1.Why does MFCA-Net provide almost no improvement with MaxViT-T? 2.How do the authors justify the lightweight claim given the large increase in GFLOPs?

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

This paper presents MFCA-Net, by augmenting an RGB detector with fixed Gaussian, Difference-of-Gaussians, and Laplacian responses computed from grayscale input. A small convolutional encoder predicts spatial softmax weights over the three responses; their projected weighted sum is scaled by a learnable parameter gamma and added to the unweighted sum of the three filter responses. The resulting feature is concatenated with RGB as a fourth input channel. The proposed method is evaluated on VisDrone-DET and DroneVehicle, showing improvement.

Critical strengths
  1. The method is straightforward to integrate into an RGB detector and does not require another sensor or additional annotations, making it a potentially useful input enhancement.
  2. The experiment provides preliminary evidence that the learned fusion is useful beyond simply supplying fixed-filter responses.
Critical weaknesses
  1. The experiments do not yet establish the value of the frequency-specific design relative to its computational cost. The main ingredients are fixed image filters and learned spatial weighting. Table 2 compares against raw RGB, individual filters, and the MFPR-only variants, but not a similarly sized learned RGB front end or learned, spatially constant fusion weights. It does not separate the benefit of the proposed frequency structure and content adaptation from that of adding a trainable input module. The Gaussian-only CAFF variant already reaches 17.6 mAP on VisDrone, versus 17.7 for the proposed Gaussian-Laplacian version, implying that the improvement is extremely minor. This matters because the overhead is small in parameters but substantial in computation. In Table 1, YOLOv8n on VisDrone increases from 4.057 to 8.665 GFLOPs (about 114%), and speed falls from 85.670 to 72.745 FPS (about 15%). The RGB-only P2 model in Table 2 achieves 19.8 mAP at 6.145 GFLOPs, compared with 17.7 at 8.665 GFLOPs for CAFF without P2, although the P2 model has slightly lower measured FPS. CAFF does improve the P2 baseline from 19.8 to 20.3; the issue is whether it offers a compelling tradeoff against other uses of the same compute.
  2. The reliability and interpretation of the reported gains remain uncertain. Section 3.2 claims that the experimental evaluation follows COCO. However, Figure 6, Module E and its caption (p. 18), describe averaging AP/mAP calculated separately for each image. Standard COCO AP instead accumulates detections across images within each category before forming the precision-recall curve. These are different metrics. This could be an error in the explanation rather than the implementation, but the paper does not identify the evaluator sufficiently to resolve it. The tables also provide no variation across training runs. This is particularly consequential for the 0.1- and 0.3-point MaxViT gains used to support the cross-backbone claim. Table 5 lists dataset split sizes, but the experimental sections do not clearly identify which evaluation partition produced Tables 1–2 or which split was used for filter selection. Training duration, training batch size, and pretraining/initialization are not clearly specified; the batch size of 1 in Section 4.2 belongs to the speed benchmark. These omissions prevent a reliable assessment of how reproducible the small improvements are.
Questions for authors
  1. Were Tables 1-2 produced using standard COCO evaluation, with detections accumulated across images, or the per-image averaging described in Figure 6? Please identify the evaluation implementation, the partitions used for the reported results, and the split used for filter and checkpoint selection. If the implementation follows Figure 6, how do the results change under standard COCO evaluation?
  2. Please specify the training duration, training batch size, and pretraining/initialization settings for each backbone. Do the baseline/CAFF gains persist across matched random seeds, particularly for MaxViT? Reporting the mean and variation of the paired mAP differences would help establish whether these small gains are reproducible.
  3. Does CAFF retain its advantage over a learned RGB front end at comparable parameter and inference budgets, and over learned spatially constant weights on the same fixed filters? These controls would help distinguish the contribution of the frequency priors and spatial adaptation from that of an additional trainable module.
  4. Equation (2) defines MFPR as a sum, whereas the following text and Figure 1 require three separate prior channels. Which representation is used in the MFPR-only rows of Table 2? For the Gaussian-only construction on p. 9, the components sum exactly to I; for the Gaussian-Laplacian construction they do not. Please distinguish concatenation from summation in the implementation and clarify whether signed filter responses are preserved.
Reviewer_QFYLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MFCA-Net to improve UAV-view object detection by incorporating multi-frequency image priors with minimal additional parameters. It uses fixed Gaussian, Difference-of-Gaussians, and Laplacian filters to extract low-, mid-, and high-frequency features, then introduces CAFF to adaptively fuse them into a single channel concatenated with RGB before the detector backbone. On VisDrone-DET, YOLOv8n+MFCA-Net improves mAP from 16.5 to 17.7 and small-object mAP from 7.7 to 8.6, with only 11,751 additional parameters, although GFLOPs increase from 4.06 to 8.67. Experiments on DroneVehicle and with a MaxViT-T backbone show smaller but consistent mAP gains, while ablations examine different frequency priors, fusion designs, and an additional P2 detection scale.

Critical strengths

The method is simple and easy to integrate into existing detectors, using fixed Gaussian, DoG, and Laplacian filters with lightweight adaptive fusion. It adds only 11,751 parameters to YOLOv8n while improving mAP by 1.2 points on VisDrone-DET, although computational cost increases substantially. The experiments cover two datasets and two different backbones, with consistent mAP improvements across all four settings. The ablations further examine individual frequency priors, adaptive fusion, and the P2 detection scale, providing evidence that the proposed design can improve detection performance across different configurations.

Critical weaknesses

MFCA-Net mainly combines conventional Gaussian, DoG, and Laplacian filters with a lightweight attention-based fusion module. These are well-established techniques, and the paper offers limited evidence that their combination provides a fundamentally new insight beyond existing frequency-based feature enhancement methods. The methodological novelty therefore appears insufficient for ICLR.

The reported gains are modest relative to the computational overhead. On VisDrone-DET, YOLOv8n improves by only 1.2 mAP while GFLOPs more than double from 4.06 to 8.67 and FPS drops from 85.67 to 72.75. With MaxViT-T, the improvement is just 0.1 mAP. These results weaken the paper's claim of an efficient and broadly effective enhancement method.

Questions for authors

How does CAFF differ in substance from existing attention-based fusion methods? A direct comparison with simple learnable weighted summation or standard attention modules would clarify whether its specific design is necessary.

Why does adding only 11,751 parameters increase YOLOv8n's GFLOPs from 4.06 to 8.67 for a 1.2-point mAP gain? Does MFCA-Net still outperform stronger baselines at comparable computational costs?

Why does the gain shrink to just 0.1 mAP with MaxViT-T on VisDrone-DET? Are these small improvements consistent across multiple random seeds, or could they reflect normal training variation?

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

Related papers

Discussion (0)

Sign in to comment.