Improving Multispectral Pedestrian Detection by Addressing Modality Imbalance Problems (ECCV 2020)

• paper download: https://arxiv.org/pdf/2008.03043.pdf
• the introduction PPT: https://github.com/CalayZhou/MBNet/blob/master/MBNet-3128.pdf
• video demo: https://www.bilibili.com/video/BV1Hi4y137aS

This code is tested on [Ubuntu18.04, tensorflow1.14, keras2.1.6, python3.6,cuda10.0,cudnn7.6].

make sure the GPU enviroment is the same as above (cuda10.0,cudnn7.6), otherwise you may have to compile the nms and utils according to https://github.com/endernewton/tf-faster-rcnn. Besides, check the keras version is keras2.1, i find there may be some mistakes if the keras version is higher. To be as simple as possible, I recommend installing the dependencies with Anaconda as follows:

1. conda create -n python36 python=3.6
2. conda activate python36
3. conda install cudatoolkit=10.0
4. conda install cudnn=7.6
5. conda install tensorflow-gpu=1.14
6. conda install keras=2.1
7. conda install opencv
8. python demo.py

• train data (download to ./data/kaist)
  Baidu Cloud(extract code: ABCD) or KAIST website
• test data (download to ./data/kaist_test)
  Baidu Cloud(extract code: ABCD) or Google Drive
• ResNet50 pretrained model (download to ./data/models/)
  Baidu Cloud(extract code: ABCD) or Google Drive
• MBNet model(download to ./data/models/)
  Baidu Cloud(extract code: ABCD) or Google Drive

1. Check the MBNet model is available at ./data/models/resnet_e7_l224.hdf5
2. Run the script: python demo.py
3. The detection result is saved at ./data/kaist_demo/.

1. Check the MBNet model is available at ./data/models/resnet_e7_l224.hdf5
2. Set weight_path , test_file , lwir_test_file in demo_video.py
3. Run the script: python demo_video.py
4. The detection result videos saved at MBNet directory.

1. check the MBNet model is available at ./data/models/resnet_e7_l224.hdf5 and the test data is available at ./data/kaist_test.
2. Run the script: python test.py
3. The test results are saved at ./data/result/.
4. open the KAISTdevkit-matlab-wrapper and run the demo_test.m.

1. Check the ResNet50 pretrained model is available at ./data/models/double_resnet.hdf5 and the train data is available at ./data/kaist.
2. Run the script: python train.py
3. The trained models are saved at ./output.
4. Evaluate model performance as above.

Please download the Matlab implemented comparison code [Baidu Cloud(extract code: ABCD) or Google Drive] and run the script according to the README.txt.

Thanks to Liu Wei, this pipeline is largely built on his ALFNet code available at: https://github.com/liuwei16/ALFNet.

If you find our work useful in your research, please consider citing:

@inproceedings{MBNet-ECCV2020,
    author = {Kailai Zhou and Linsen Chen and Xun Cao},
    title = {Improving Multispectral Pedestrian Detection by Addressing Modality Imbalance Problems},
    booktitle = ECCV,
    year = {2020}
}