GeoViG: Geometry-Aware Graph Reasoning for Mobile Vision Tasks

Model Variants

🖼️ Image Classification — ImageNet-1K

Model	Params (M)	MACs (G)	Top-1 Acc (%)	Checkpoint (pth)
GeoViG-Ti	3.5	0.9	75.2	Download
GeoViG-S	5.0	1.2	77.5	Download
GeoViG-M	10.3	2.2	80.7	Download
GeoViG-B	19.7	4.5	82.4	Download

Trained for 300 epochs on 8× NVIDIA A100 GPUs, batch size 1024, AdamW optimizer (lr=5e-4, weight decay=0.05), cosine schedule with 5-epoch warmup. Augmentations: RandAugment, Mixup (p=0.8), CutMix (p=1.0).

For CoreML checkpoint, please check: https://huggingface.co/OmarAlasqa/GeoViG/tree/main/CoreML For IPA chechpoints, please check: https://huggingface.co/OmarAlasqa/GeoViG/tree/main/IPA

📦 Object Detection & Instance Segmentation — MS COCO 2017

Backbone used with Mask R-CNN, 1× schedule (12 epochs), pretrained on ImageNet-1K.

Backbone	Params (M)	Box AP	Box AP₅₀	Box AP₇₅	Mask AP	Mask AP₅₀	Mask AP₇₅	Checkpoint
GeoViG-M	10.3	40.7	62.4	44.1	37.7	59.6	40.5	Download
GeoViG-B	19.7	42.5	64.0	46.8	38.9	61.2	41.7	Download

🏥 Medical Image Segmentation

Kvasir-SEG — Polyp Segmentation

Model	Params (M)	mAP	Dice ↑	IoU ↑	Hausdorff Dist ↓	Checkpoint
GeoViG-M	29.57	0.990	0.945	0.909	12.94	Download

Data Science Bowl 2018 — Nuclei Segmentation

Model	Params (M)	mAP	Dice ↑	IoU ↑	Hausdorff Dist ↓	Checkpoint
GeoViG-M	29.57	0.859	0.908	0.839	5.19	Download

🚀 Usage

Please check the github repo: https://github.com/OmarAlsaqa/GeoViG

Citation

If you use GeoViG in your research, please cite:

@article{alsaqa2026geovig,
  title     = {GeoViG: Geometry-Aware Graph Reasoning for Mobile Vision Tasks in Natural and Medical Images},
  author    = {Alsaqa, Omar and Mohammed, Emad and Aleem, Saiqa},
  journal   = {Under Review at IEEE EMBC},
  year      = {2026}
}

Acknowledgements

This work builds upon MobileViG and uses the MMDetection framework for detection and segmentation experiments. Training was performed on the Compute Canada A100 cluster.

OmarAlasqa
/

GeoViG