Awesome

GNN-CCA

This is a DGL implementation of GNN-CCA for multi-view detections.

The original GNN-CCA was implemented in PyGeometric. This repo re-implements in DGL. Both are using PyTorch.

cover

Installation

Install PyTorch (>= 1.9.0) and DGL (Python 3.7)
Install other requirements and warmup_scheduler:

pip install -r requirements.txt

pip install git+https://github.com/ildoonet/pytorch-gradual-warmup-lr.git

Install torchreid (follow its instruction)

Preparation

Dataset

Assume DATA_NAME is the directory in dataset folder.

Download dataset. Please refer to dataset/${DATA_NAME}/README.md
Run python preprocess/${DATA_NAME}.py with ${DATA_NAME} be the lower case. For example:

python preprocess/epfl.py

Model

Download ReID model from here and assume its path is PATH_TO_REID_MODEL.
You can also download the trained model from table below and assume its path is PATH_TO_MODEL.

Model Name	Download Link	Train Dataset	Test Dataset	Performance
gnn_cca_epfl	Google Drive <br/> Baidu Drive (m1he)	EPFL	EPFL	See Table 1

Note: all the trainings are under the same default configuration.

Table 1:

Test Sequence	ARI	AMI	H	C	V-m	ACC
basketball	72.998	58.406	52.270	66.799	58.476	97.035
laboratory	91.983	86.821	85.714	88.184	86.880	98.268
passageway	71.211	66.528	64.718	80.970	67.081	93.573
terrace	90.749	84.478	83.060	86.228	84.526	98.300

Training

Training model on a specific dataset.

For example, training on EPFL dataset with all sequences:

python run.py --train --reid-path ${PATH_TO_REID_MODEL} --epfl --seq-name all

training on EPFL dataset with specific sequences:

python run.py --train --reid-path ${PATH_TO_REID_MODEL} --epfl --seq-name terrace passageway

You can also change the ReID model (served as the feature extractor) refer to here, and assume its name is NAME_OF_REID_MODEL. Then you can train your model by running:

python run.py --reid-name ${NAME_OF_REID_MODEL} --reid-path ${PATH_TO_REID_MODEL} ...

After each epoch, the trained model will be saved in the directory assigned by --output.

Finally, denote PATH_TO_MODEL as the trained model.

Testing

Testing model on a specific dataset.

For example, testing on EPFL dataset with all sequences:

python run.py --test --reid-path ${PATH_TO_REID_MODEL} --ckpt ${PATH_TO_MODEL} --epfl --seq-name all

and you can also plot the results of some sequences by adding --visualize:

python run.py --test --reid-path ${PATH_TO_REID_MODEL} --ckpt ${PATH_TO_MODEL} --epfl --seq-name terrace laboratory --visualize

The results will be saved in the directory assigned by --output.

Results

Result of EPFL-basketball:

result1

Result of EPFL-laboratory:

result2

Citation

@article{luna2022gnncca,
  title={Graph Neural Networks for Cross-Camera Data Association},
  author={Luna, Elena and SanMiguel, Juan C. and Martínez, José M. and Carballeira, Pablo},
  journal={arXiv preprint arXiv:2201.06311},
  year={2022}
}