Skip to content

Latest commit

 

History

History
35 lines (32 loc) · 2.27 KB

run.md

File metadata and controls

35 lines (32 loc) · 2.27 KB

AlphaPose Usage & Examples

Here, we first list the flags of this script and then give some examples.

Flags

  • --indir: Directory of the input images. All the images in the directory will be processed.
  • --list: A text file list for the input images
  • --video: Read video and process the video frame by frame.
  • --outdir: Output directory to store the pose estimation results.
  • --vis: If turned-on, it will render the results and visualize them.
  • --save_img: If turned-on, it will render the results and save them as images in $outdir/vis.
  • --save_video: If turned-on, it will render the results and save them as a video.
  • --vis_fast: If turned on, it will use faster rendering method. Default is false.
  • --format: The format of the saved results. By default, it will save the output in COCO-like format. Alternative options are 'cmu' and 'open', which saves the results in the format of CMU-Pose or OpenPose. For more details, see output.md
  • --conf: Confidence threshold for human detection. Lower the value can improve the final accuracy but decrease the speed. Default is 0.1.
  • --nms: Confidence threshold for human detection. Increase the value can improve the final accuracy but decrease the speed. Default is 0.6.
  • --detbatch: Batch size for the detection network.
  • --posebatch: Maximum batch size for the pose estimation network. If you met OOM problem, decrease this value until it fit in the memory.
  • --sp: Run the program using a single process. Windows users need to turn this flag on.
  • --inp_dim: The input size of detection network. The inp_dim should be multiple of 32. Default is 608.

Examples

  • Run AlphaPose for all images in a folder ,save the results in the format of CMU-Pose and save the images:
python3 demo.py --indir examples/demo/ --outdir examples/results/ --save_img --format cmu
  • Run AlphaPose for a video, save the video and use faster rendering method:
python3 video_demo.py --video examples/input.mp4 --outdir examples/results/ --save_video  --vis_fast
  • Run AlphaPose for a video, speeding up by increasing the confidence and lowering the NMS threshold.:
python3 video_demo.py --video examples/input.mp4 --outdir examples/results/  --conf 0.5 --nms 0.45