zimenglan-sysu-512 / mscnn

A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection

by Zhaowei Cai, Quanfu Fan, Rogerio Feris and Nuno Vasconcelos

This implementation is written by Zhaowei Cai at UC San Diego.

Introduction

MS-CNN is a unified multi-scale object detection framework based on deep convolutional networks, which includes an object proposal sub-network and an object detection sub-network. The unified network can be trained altogether end-to-end.

Citations

If you use our code/model/data, please cite our paper:

@inproceedings{cai16mscnn,
  author = author = {Zhaowei Cai and Quanfu Fan and Rogerio Feris and Nuno Vasconcelos},
  Title = {A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection},
  booktitle = {ECCV},
  Year  = {2016}
}

Requirements

1. cuDNN is required to have the same running speed described in our paper. For now, only cuDNN 7.0 is supported.

2. If you want to use our MATLAB scripts to run the detection demo, caffe MATLAB wrapper is required. Please build matcaffe before running the detection demo.

3. This code has been tested on Ubuntu 14.04 with an NVIDIA Titan GPU.

Installation

1. Clone the MS-CNN repository, and we'll call the directory that you cloned MS-CNN into MSCNN_ROOT

   git clone https://github.com/zhaoweicai/mscnn.git

2. Build MS-CNN

   cd $MSCNN_ROOT/
   # Follow the Caffe installation instructions here:
   #   http://caffe.berkeleyvision.org/installation.html
   
   # If you're experienced with Caffe and have all of the requirements installed
   # and your Makefile.config in place, then simply do:
   make all -j 16
   
   # If you want to use MSCNN detection demo, build MATLAB wrapper as well
   make matcaffe

Training MS-CNN (KITTI car)

1. Set up KITTI dataset by yourself.

2. Get the training data for KITTI

   cd $MSCNN_ROOT/data/
   sh get_kitti_data.sh

   This will download train/val split image lists for the experiments, and window files for training/finetuning MS-CNN models. You can also use the provided MATLAB scripts mscnn_kitti_car_window_file.m under $MSCNN_ROOT/data/kitti/ to generate your own window files. If you use the provided window files, replace /your/KITTI/path/ in the files to your KITTI path.

3. Download VGG16 from Caffe Model Zoo, and put it into $MSCNN_ROOT/models/VGG/.

4. Now you can start to train MS-CNN models. Multiple shell scripts are provided to train different models described in our paper. We take mscnn-7s-576-2x for example.

   cd $MSCNN_ROOT/examples/kitti_car/mscnn-7s-576-2x/
   sh train_mscnn.sh

   As described in the paper, the training process is split into two steps. Usually the first step can be shared by different models if you only have modifications on detection sub-network. For example, the first training step can be shared by mscnn-7s-576-2x and mscnn-7s-576. Meanwhile, log files will be generated along the training procedures.

Pretrained model (KITTI car)

Download pre-trained MS-CNN models

cd $MSCNN_ROOT/examples/kitti_car/
sh fetch_mscnn_car_model.sh

This will download the pretrained model for KITTI car into $MSCNN_ROOT/examples/kitti_car/mscnn-8s-768-trainval-pretrained/. You can produce exactly the same results as described in our paper with these pretrained models.

Testing Demo (KITTI car)

Once the pretrained models or models trained by yourself are available, you can use the MATLAB script run_mscnn_detection.m under $MSCNN_ROOT/examples/kitti_car/ to obtain the detection and proposal results. Set the right dataset path and choose the model that you want to test in the demo script. The default setting is to test the pretrained model. The final results will be saved as .txt files.

Disclaimer

1. This implementation is developed on an old version of Caffe. Some new layers of current Caffe may not be supported in this implementation. But some layers (e.g. "batch_norm", "scale", "bias") are merged into this repository, such that ResNet is supported.

2. The CPU version is not fully tested. The GPU version is strongly recommended.

3. Since some changes have been made after ECCV submission, you may not have exactly the same results in the paper by training your own models. But you should have equivelant performance.

4. Since the numbers of training samples vary vastly for different classes, the model robustness varies too (car>ped>cyc).

5. Although the final results we submitted were from model mscnn-8s-768-trainval, our later experiments have shown that mscnn-7s-576-2x-trainval can achieve even better performance for car. With this model, detection can run 4fps. For ped/cyc however, the performance decreases due to the much less training instances.

6. If the training does not converge or the performance is very bad, try some other random seeds. You should obtain fair performance after a few tries. Due to the randomness, you cann't fully reproduce the same models, but the performance should be close.

If you encounter any issue when using our code or model, please let me know.