wanglimin / action.sr_cnn

code for Two-Stream SR-CNNs for Action Recognition in Videos 2016

BMVC2016 Two-Stream SR-CNNs for Action Recognition in Videos

---

Prerequisites

Caffe

• clone and build caffe from here. This caffe version is based on Limin Wang's fork [1] contains merge_batch and weighted_sum layer. In addition it exposed some protected caffe functions in the matlab interface to emulate iter_size in matlab.
• modify caffe_mex.m to the corresponding caffe matlab interface directory

Optical Flow

• extract optical flow with Limin's flow extractor

Bounding Boxes

• We extracted 118 objects' bounding boxes in all video frames using Faster-RCNN [2] (retraining is required) and obtained filtered bounding boxes taking consideration of temporal coherency and motion saliency.
• The extracted and processed bounding boxes for ucf-101 can be downloaded here. Place the downloaded mat files under imdb/cache.
• If you wish to extract the bounding boxes yourself, you need to be able to run Ren Shaoqing's Faster-RCNN (most codes are migrated into this repository with minor modifications and more comments)
  • First generate raw object detection using faster_rcnn_{dataset}.m
  • Then use action/prepare_rois_context.m to process bounding boxes as described in the paper.

---

Test

datasets

create dataset.mat using imdb/get_{name}_dataset.m (Directories may need to be adjusted!) An example of generated ucf_dataset.mat

models

• models/srcnn/{stream} contains model prototxt files

• model weights can be downloaded in the following links

  Stream	person+scene (the final proposed model in the paper)
  spatial	split1 split2 split3
  flow	split1 split2 split3

• the reported two-stream results in the paper are yielded from summing spatial and temporal classification scores using weight 1 : 3.

• other models mentioned in the paper experiments can be provided if the demand is large.

run

in matlab

% test spatial
test_spatial('model_path', path_to_weights, 'split', 1)

% test flow
`test_flow('model_path', path_to_weights, 'split', 1)`

---

Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems (pp. 91-99).

Wang, L., Xiong, Y., Wang, Z., & Qiao, Y. (2015). Towards good practices for very deep two-stream convnets. arXiv preprint arXiv:1507.02159.