pengxj / action-faster-rcnn

action-faster-rcnn

This repository is a strongly modified version for action detection originally from py-faster-rnn for my ECCV16 paper. It wraps three popular action detection dataset classes: UCF-Sports, JHMDB, and UCF101. Also, it provides useful action detection evaluation scripts for both frame level and video level. Note the results on UCF101 are updated at https://hal.inria.fr/hal-01349107/file/eccv16-pxj-v3.pdf dut to some annotation parsing errors.

Installation

1. Clone this reporsitory

git clone --recursive https://github.com/pengxj/action-faster-rcnn.git

2. Build the Cython modules which mainly compiles the nms module

cd $THIS_ROOT/lib
make

3. Build Caffe and pycaffe

cd $THIS_ROOT/caffe-fast-rcnn-faster-rcnn-upstream-33f2445
# Now follow the Caffe installation instructions here:
#   http://caffe.berkeleyvision.org/installation.html

4. Dive into the code

dataset classes: lib/datasets/ucfsports.py JHMDB.py UCF101.py
training script: action_experiments/scripts/train_action_det.sh
evaluation scripts: action_tools/action_util.py ucfsports_eval.py jhmdb_eval.py ucf101_eval.py fusion_eval.py eval_linked_results.py
script for merging 2 stream models: action_tools/net_surgery_rgbflow.py

Run experiments

The entire pipeline for two-stream rcnn includes optical flow extraction, r-cnn training, frame-level detecting, linking and evaluation. All these are included in this repository.

If you just want to get the final video AP, you download the UCF101 linked results and run the eval_linked_results script. The folder 'action_results' includes linked results for UCF-Sports and JHMDB datasets.

video mAP results with different iou thresholds (without multi-region scheme):

	0.2	0.5	0.75	0.5:0.95
UCF-Sports	95.12	95.12	47.33	50.95
JHMDB	72.75	72.11	48.15	42.23
UCF101 Split 1	73.20	35.91	1.55	8.76

python action_tools/eval_linked_results.py --imdb UCF101_RGB_1_FLOW_5_split_0 --res path/to/ucf101_vdets_3scales_rgb1flow5.pkl

{0.05: 0.7881, 0.1: 0.7745, 0.2: 0.7320, 0.3: 0.6630, 0.4: 0.5604, 0.5: 0.3591, 0.6: 0.1469, 0.7: 0.0349}

python action_tools/eval_linked_results.py --imdb JHMDB_RGB_1_FLOW_5_split_2 --res action_results/jhmdb_s03_vdets_3scales_rgb1flow5.pkl

{0.5: 0.7124, 0.4: 0.7124, 0.2: 0.7139, 0.05: 0.7139, 0.6: 0.7028, 0.3: 0.7134, 0.1: 0.7139, 0.7: 0.6009}

python action_tools/eval_linked_results.py --imdb JHMDB_RGB_1_FLOW_5_split_1 --res action_results/jhmdb_s02_vdets_3scales_rgb1flow5.pkl

{0.5: 0.7304, 0.4: 0.7360, 0.2: 0.7412, 0.05: 0.7414, 0.6: 0.7063, 0.3: 0.7412, 0.1: 0.7414, 0.7: 0.6004}

python action_tools/eval_linked_results.py --imdb JHMDB_RGB_1_FLOW_5_split_0 --res action_results/jhmdb_s01_vdets_3scales_rgb1flow5.pkl

{0.5: 0.7207, 0.4: 0.7240, 0.2: 0.7273, 0.05: 0.7299, 0.6: 0.6909, 0.3: 0.7244, 0.1: 0.7273, 0.7: 0.5974}

python action_tools/eval_linked_results.py --imdb UCF-Sports_RGB_1_FLOW_5_split_0 --res action_results/ucfsports_vdets_3scales_rgb1flow5.pkl

{0.5: 0.9512, 0.4: 0.9512, 0.2: 0.9512, 0.05: 0.9512, 0.6: 0.9034, 0.3: 0.9512, 0.1: 0.9512, 0.7: 0.7370}

And for the 'imdb' option, you can find them in dir action_experiments/listfiles/ which are actually the names of files.

Citation

If you find this repository useful in your research, please consider citing:

@inproceedings{peng2016multi,
title={Multi-region two-stream R-CNN for action detection},
author={Peng, Xiaojiang and Schmid, Cordelia},
booktitle={European Conference on Computer Vision},
pages={744--759},
year={2016},
organization={Springer}}


@inproceedings{renNIPS15fasterrcnn,
    Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
    Title = {Faster {R-CNN}: Towards Real-Time Object Detection
             with Region Proposal Networks},
    Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
    Year = {2015}
}