FCBNet: A Parameter-efficient Convolutional Approach for Camouflaged Weed Detection in Multispectral Aerial Imagery

Leo Thomas Ramos, Angel D. Sappa
Correspondence: ltramos@cvc.uab.cat

Computer Vision Center (CVC), Universitat Autònoma de Barcelona (UAB), Spain

Announcements

• Official implementation and model code are available
• FCBNet paper is now available here
• FCBNet has been accepted at the 2nd Workshop on Subtle Visual Computing (SVC), CVPR 2026

About the project

FCBNet is a parameter-efficient semantic segmentation model for camouflaged weed detection in aerial imagery. It combines a fully frozen ConvNeXt encoder with lightweight Feature Correction Blocks (FCB) and an FPN-style decoder. The main design goal is to keep strong segmentation performance while drastically reducing trainable parameters and training cost.

Installation

pip install -r requirements.txt

or install directly from source:

pip install .

Quick usage

FCBNet can be instantiated directly from fcbnet.py and used within any training workflow as follows:

from fcbnet import FCBNet

model = FCBNet(
    in_channels=3,
    num_classes=4,
)

The default arguments are already configured with the best-performing setup from our experiments. It can still be tuned through parameters such as variant, pretrained, freeze_backbone, fpn_dim, and the FCB-related settings if needed.

Input and output

• Input: tensor of shape (B, C, H, W)
• Output: logits of shape (B, num_classes, H, W)

Example forward pass:

import torch
from fcbnet import FCBNet

model = FCBNet(in_channels=3, num_classes=4)
x = torch.randn(2, 3, 256, 256)
y = model(x)
print(y.shape)  # (2, 4, 256, 256)

Datasets

To evaluate FCBNet, we use two camouflaged weed detection/segmentation datasets based on aerial imagery:

• WeedBananaCOD
• WeedMap

Main results

WeedBananaCOD

Model	Modality	IoU_BG	IoU_W	mIoU	Inference time (s)	Training time (h)	Total params (M)	Trainable params (M)	GFLOPS
U-Net	RGB	0.990	0.715	0.851	0.0243	0.097	32.521	32.521	42.771
U-Net	RGB-NIR	0.990	0.720	0.855	0.0257	0.100	32.524	32.524	42.976
DeepLabV3+	RGB	0.988	0.721	0.855	0.0270	0.123	26.678	26.678	36.762
DeepLabV3+	RGB-NIR	0.991	0.725	0.858	0.0390	0.125	26.681	26.681	36.968
SegFormer	RGB	0.989	0.695	0.842	0.0203	0.310	84.595	84.595	49.733
SegFormer	RGB-NIR	0.990	0.738	0.864	0.0295	0.312	84.598	84.598	49.784
FCBNet-tiny	RGB	0.990	0.717	0.853	0.0139	0.060	30.604	2.015	13.151
FCBNet-tiny	RGB-NIR	0.991	0.740	0.865	0.0149	0.062	30.605	2.015	13.176
FCBNet-large	RGB	0.991	0.746	0.868	0.0360	0.169	202.322	4.555	19.504
FCBNet-large	RGB-NIR	0.992	0.771	0.881	0.0381	0.171	202.325	4.555	19.554

WeedMap

Model	Modality	IoU_BG	IoU_W	mIoU	Inference time (s)	Training time (h)	Total params (M)	Trainable params (M)	GFLOPS
U-Net	RGB	0.975	0.427	0.701	0.0109	0.119	32.521	32.521	30.073
U-Net	RGB-NIR	0.981	0.508	0.745	0.0112	0.126	32.524	32.524	30.218
U-Net	RGB-NIR-RE	0.986	0.550	0.768	0.0117	0.131	32.528	32.528	30.362
DeepLabV3+	RGB	0.983	0.509	0.746	0.0190	0.161	26.678	26.678	25.849
DeepLabV3+	RGB-NIR	0.984	0.515	0.749	0.0190	0.170	26.681	26.681	25.993
DeepLabV3+	RGB-NIR-RE	0.986	0.534	0.760	0.0200	0.189	26.684	26.684	26.137
SegFormer	RGB	0.986	0.506	0.746	0.0177	0.420	84.595	84.595	34.116
SegFormer	RGB-NIR	0.985	0.532	0.758	0.0180	0.421	84.598	84.598	34.153
SegFormer	RGB-NIR-RE	0.985	0.535	0.760	0.0189	0.421	84.601	84.601	34.189
FCBNet-tiny	RGB	0.985	0.537	0.761	0.0095	0.083	30.604	2.015	9.247
FCBNet-tiny	RGB-NIR	0.986	0.540	0.763	0.0103	0.091	30.605	2.015	9.264
FCBNet-tiny	RGB-NIR-RE	0.986	0.543	0.764	0.0107	0.098	30.607	2.015	9.282
FCBNet-large	RGB	0.987	0.546	0.766	0.0252	0.201	202.322	4.555	13.714
FCBNet-large	RGB-NIR	0.986	0.548	0.767	0.0251	0.211	202.325	4.555	13.749
FCBNet-large	RGB-NIR-RE	0.987	0.551	0.769	0.0257	0.215	202.329	4.555	13.784

Why FCBNet is efficient

• Fully frozen ConvNeXt backbone
• Lightweight Feature Correction Blocks (pointwise + depthwise convolutions)
• Compact FPN decoder and segmentation head
• Lower memory footprint and faster optimization

Citation

If you find this work useful, please star ⭐️⭐️⭐️ our repository and cite our paper:

@InProceedings{Ramos_2026_CVPR,
    author    = {Ramos, Leo Thomas and Sappa, Angel D.},
    title     = {A Parameter-efficient Convolutional Approach for Camouflaged Weed Detection in Multispectral Aerial Imagery},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2026},
    pages     = {1349-1358}
}

License

Distributed under the GNU General Public License v3.0. See LICENSE for more information.