zhangxujinsh / tpn

Tubelet Proposal Network

TPN: Tubelet Proposal Network

Introduction

TPN, short for Tubelet Proposal Network, is a deep learning framework for video object detection, originally designed for ImageNet VID dataset.

This framework mainly contains two components shown below:

The Tubelet Proposal Network generates hundreds of tubelet proposals based on static image box proposals, and the Encoder-decoder LSTM encodes the tubelet visual features into memory and decodes the information to classify each tubelet box proposal into different classes. More details in the paper in CVPR 2017.

Citation

If you are using the TPN code in your project, please cite the following publication.

@inproceedings{kang2017tpn,
  title={Object Detection in Videos with Tubelet Proposal Networks},
  author={Kang, Kai and Li, Hongsheng and Xiao, Tong and Ouyang, Wanli and Yan, Junjie and Liu, Xihui and Wang, Xiaogang},
  Booktitle = {CVPR},
  Year = {2017}
}

Installations

Dependencies

1. Faster R-CNN
2. vdetlib
3. Caffe with MPI
4. TensorFlow v0.8

Instructions

>> # clone this repository
>> git clone --recursive [email protected]:myfavouritekk/TPN.git
>> # compile external dependencies
>> cd $TPNROOT/external/vdetlib && make
>> cd $TPNROOT/external/py-faster-rcnn/lib/ && make
>> # compile caffe with MPI
>> cd $TPNROOT/external/caffe-mpi
>> mkdir build && cd build
>> cmake .. && make -j

Usage

1. Generate .vid file

>> python TPN/external/vdetlib/tools/gen_vid_proto_file.py vid_name root_dir out_file

2. Generate static proposals
   1. static proposals for ImageNet VID generated with ResNet RPN.
   2. Generating the .box file for region proposals proposal files contains two fields: boxes contains box proposals, images contains image frame names. The images can be the following forms: video_name/frame_name, subset/video_name/frame_name.

>> python tools/data/box_proto_from_proposals.py proposal_file vid_root save_dir

3. Generating tubelets with TPN:
   1. Download CVPR 2017 models and extract to CVPR2017_models.
   2. generating tubelets with multiple GPUs:

>> cat video_list.txt | parallel -j 8 python tools/propagate/sequence_roi_propagation.py /path/to/vids/{}.vid /path/to/boxes/{}.box /path/to/save/track/dir --def CVPR2017_models/tpn/googlenet/deploy.prototxt --param CVPR2017_models/tpn/googlenet/hkbn_4d_fast_rcnn_det2_vid1_resnet_rpn_sequence_roi_length_5_iter_120000.caffemodel --bbox_mean CVPR2017_models/tpn/googlenet/vid_2015_train_20_length_5_gt_mean.pkl --bbox_std CVPR2017_models/tpn/googlenet/vid_2015_train_20_length_5_gt_std.pkl --bbox_pred_layer bbox_pred_vid_pair_roi --length 20 --sample_rate 20 --cfg experiments/cfgs/resnet_rpn.yml --num_per_batch 150 --job {%} --gpus {0..7}

4. Detection:

>> python TPN/tools/propagate/track_propagation.py vid_file track_file save_file --def CVPR2017_models/vid/depoly.prototxt --net CVPR2017_models/vid/hkbn_4d_fast_rcnn_vid_only_iter_90000.caffemodel --crf TPN/experiments/cfgs/craft.yml --num_per_batch 300 --bbox_mean CVPR2017_models/vid/vid_2015_train_20_bbox_means.pkl --bbox_std CVPR2017_models/vid/vid_2015_train_20_bbox_stds.pkl

Beyond demo

1. Generating the .vid file for a video

>> python external/vdetlib/tools/gen_vid_proto_file.py vid_name root_dir out_file.vid

2. Generating the .box file for region proposals proposal files contains two fields: boxes contains box proposals, images contains image frame names. The images can be the following forms: video_name/frame_name, subset/video_name/frame_name.

>> python tools/data/box_proto_from_proposals.py proposal_file vid_root save_dir

Models

1. TPN models in CVPR2017_models
2. Detection models in CVPR2017_models

Links

1. PDF
2. Poster
3. Demo video

License

TPN is released under the MIT License.