MVDGC detects people jointly in bird's-eye-view (BEV) ground space and per-camera image space using a set of learnable 3D cylindrical queries. Each query is initialized across the BEV plane, projected into every camera view to sample multi-view features via sparse deformable attention, and iteratively refined to produce accurate ground-plane positions and per-view bounding boxes — all in a single forward pass, with no depth estimation or explicit 3D reconstruction.

git clone <repo-url>
cd MVDGC

conda create -n mvcyde python=3.8 -y
conda activate mvcyde

pip install torch==2.1.0 torchvision==0.16.0 \
    --index-url https://download.pytorch.org/whl/cu121

pip install mmcv-full==1.7.2 -f \
    https://download.openmmlab.com/mmcv/dist/cu121/torch2.1/index.html
pip install mmengine==0.10.7

Note: mmcv-full must match your CUDA and PyTorch versions exactly. If the prebuilt wheel is not available, build from source: MMCV_WITH_OPS=1 pip install mmcv-full==1.7.2

pip install -r requirements.txt

The multi-scale deformable attention CUDA kernel must be compiled before training.

cd models/ops
bash make.sh
cd ../..

Verify the build succeeded:

python models/ops/test.py

Set the ROOT path in the config file for your chosen dataset:

# configs/wildtrack.yaml
DATASET:
  ROOT: '/path/to/Wildtrack_dataset'

# configs/multiviewx.yaml
DATASET:
  ROOT: '/path/to/MultiviewX'

# configs/gmvd.yaml
DATASET:
  ROOT: '/path/to/GMVD'

# Wildtrack
python main_sparse_deform_batch.py \
    --cfg ./configs/wildtrack.yaml \
    --exp_name wildtrack_run1

# MultiviewX
python main_sparse_deform_batch.py \
    --cfg ./configs/multiviewx.yaml \
    --exp_name multiviewx_run1

# GMVD
python main_sparse_deform_batch.py \
    --cfg ./configs/gmvd.yaml \
    --exp_name gmvd_run1

Set the checkpoint name inside test_sparse_deform.py:

# test_sparse_deform.py  (line ~80)
weight_name = "wildtrack_epoch_28.pth"   # file must be inside output/

Then run:

python test_sparse_deform.py --cfg ./configs/wildtrack.yaml

If you use this work in your research, please cite:

@article{mvdgc2026,
  title   = {MVDGC: Joint 3D and 2D Multi-view Pedestrian Detection via Dual Geometric Constraints},
  author  = {},
  year    = {2024}
}

This project builds upon:

• Deformable DETR — sparse deformable attention
• mmdetection — ViT backbone and SFP neck
• MVDeTr — multi-view detection framework

Please refer to their respective licenses when using this code.