Skip to content

Compute dense optical flow using TV-L1 algorithm with NVIDIA GPU acceleration.

Notifications You must be signed in to change notification settings

wizyoung/Optical-Flow-GPU-Docker

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 

Repository files navigation

Optical-Flow-GPU

Calculate dense optical flow using TV-L1 algorithm with NVIDIA GPU acceleration. The CPU version is also included. Dockerhub link

Docker image environment: OpenCV 2.4, CUDA 8, cuDNN 5.

The dense flow C++ source code for building is from yjxiong/dense_flow. The docker image is based on willprice/opencv2-cuda8. If you want to build opencv with cuda support yourself, you can refer to this dockerfile.

Requirements: docker, nvidia-docker. If you have libseccomp2 version conflict problems when installing docker on ubuntu, you can refer to this solution.

Usage

1. Directly use the prebuilt binaries (no multi gpu support)

extract_cpu [OPTION] ...  # using cpu
extract_gpu [OPTION] ...  # using gpu
extract_warp_gpu [OPTION] ...  # using gpu to extract warp flow

Avaliable options:

  • -f: video path.
  • -x: filename of flow x component.
  • -y: filename of flow y component.
  • -i: filename of extracted RGB image.
  • -b: boundary to clip the flow value. For example, -b 20 means clip the flow value beyond [-20, 20] and maps the [-20, 20] interval to [0, 255] in grayscale image space.
  • -t: flow calculation method. 0: Brox, 1: TVL1.
  • -s: step for frame sampling. 1 means no skipping frame in calculation.
  • -d: gpu id.
  • -w, -h: resize the image before flow calculation. w = resized_width, h = resized_height. 0 means no resize. Note: If you want to resize the image, w and h must be both specified to take effect.
  • -o: output format. 'dir' means saving the flow image data in directory format. 'zip' means saving the flow image data in zipped file format.

Example:

# first mount the folder containing videos to /data
nvidia-docker run -it -v path_to_mount:/data wizyoung/optical-flow-gpu bash 
mkdir /data/result
# RGB image data is abandoned here
extract_gpu -f /data/video.mp4 -x /data/result/flow_x -y /data/result/flow_y -i /dev/null -b 20 -t 1 -s 1 -d 0 -w 100 -h 100 -o zip

Result:

# path: /data/
.
├── result
│   ├── flow_x.zip
│   └── flow_y.zip
└── video.mp4

2. Batch processing using python script with multi gpu support

I included the python wrapper script with multi gpu support in the /src path:

root@eca86f630747:/src# python multi_gpu_extract.py -h
usage: multi_gpu_extract.py [-h] [--flow_type {tvl1,warp_tvl1}]
                            [--out_fmt {dir,zip}] [--num_gpu NUM_GPU]
                            [--step STEP] [--keep_frames KEEP_FRAMES]
                            [--width WIDTH] [--height HEIGHT] [--log LOG]
                            vid_txt_path out_dir

Extract optical flows with multi-gpu support.

positional arguments:
  vid_txt_path          Input txt files containing video paths.
  out_dir               Destination directory to store flow results.

optional arguments:
  -h, --help            show this help message and exit
  --flow_type {tvl1,warp_tvl1}
                        Optical flow type. Default: tvl1
  --out_fmt {dir,zip}   Output file format. Default: zip
  --num_gpu NUM_GPU     Number of GPUs. Default: 4
  --step STEP           Specify the step for frame sampling. Default: 1
  --keep_frames KEEP_FRAMES
                        Whether to save RGB frame data. Default: False
  --width WIDTH         Resize image width. Default: 0 (no resize)
  --height HEIGHT       Resize image height. Default: 0 (no resize)
  --log LOG             Output log file path. Default: ./out.log

NOTE:

(1) vid_txt_path should be a txt file each line containing a video path.

Example:

# NOTE: /data/ is the mounting point rather than the actual path
# vid_txt_path: /data/video_list.txt
/data/videos/1.mp4
/data/videos/2.mp4
/data/videos/3.mp4
...

(2) num_gpu means using gpus from 0 to num_gpu - 1.

(3) The log file records the processing progress and error (likely corrupted videos) during processing.

Example:

python multi_gpu_extract.py /data/video_list.txt /data/results --flow_type tvl1 --out_fmt zip --num_gpu 4 --step 1 --keep_frames True --width 100 --height 100 --log /data/log.log

About

Compute dense optical flow using TV-L1 algorithm with NVIDIA GPU acceleration.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published