Skip to content

Latest commit

 

History

History
76 lines (49 loc) · 8.67 KB

README.md

File metadata and controls

76 lines (49 loc) · 8.67 KB
Raw

Experimental Models

Contents

  1. Introduction
  2. Single-Network Whole-Body Pose Estimation Model
  3. Official BODY_25 Model
  4. BODY_25B Model - Option 1 (Maximum Accuracy, Less Speed))
  5. BODY_25B Model - Option 2 (Recommended))

Introduction

Very important: All of these models require OpenPose with Nvidia GPU support, they do not work on CPU or OpenCL modes.

The experimental_models directory contains our experimental models, including the whole-body model from Single-Network Whole-Body Pose Estimation, as well as instructions to make it run inside OpenPose.

Which model should I use?

  1. If you want the model that trains body, face, hands, and feet in the same network, go for BODY_135.
  2. If you want the maximum accuracy, regardless of the speed, train BODY_25B Model - Option 1.
  3. If you want to maintain the current OpenPose speed, while increasing its accuracy, go for BODY_25B Model - Option 2.
  4. If you are OK losing a bit of accuracy but simply want to fine-tune the current OpenPose, go for BODY_25.

Single-Network Whole-Body Pose Estimation Model

To use the BODY_135 model trained for the Single-Network Whole-Body Pose Estimation paper (i.e., higher accuracy but slower speed than the default BODY_25 model):

  1. Download the Caffe model from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/100_135AlmostSameBatchAllGPUs/body_135/pose_iter_XXXXXX.caffemodel into experimental_models/100_135AlmostSameBatchAllGPUs/body_135/ as pose_iter_XXXXXX.caffemodel.
  2. Copy the body_135 folder into the OpenPose models folder, such as you end up with the previous 2 files in the following paths:
    1. openpose/models/pose/body_135/pose_deploy.prototxt
    2. openpose/models/pose/body_135/pose_iter_XXXXXX.caffemodel. Note: This model corresponds to pose_iter_464000.caffemodel (i.e., trained for 464k iterations), and it was renamed simply to follow the OpenPose format.
  3. Run the official OpenPose with your usual flags, while adding the --model_pose BODY_135 --net_resolution -1x480.
  4. Alternatively to 3, for multi-scale (even higher accuracy), then use --model_pose BODY_135 --net_resolution -1712x960 --scale_number 4 --scale_gap 0.25.

In addition, we also include all the files generated for and during training in experimental_models/100_135AlmostSameBatchAllGPUs/training_results/. If you want to read the training_log.txt file, you can download it from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/100_135AlmostSameBatchAllGPUs/body_135/training_log.txt.

Official BODY_25 Model

This model is already included with the official OpenPose, so simply run the default version of OpenPose in order to use it.

To allow fine-tuning the official OpenPose model, and although this model is not "experimental", we also include all the files generated for and during its training in experimental_models/body_25/training_results/. If you want to read the training_log.txt file, you can download it from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/body_25/training_log.txt.

Note: For this model, you must use openpose/models/pose/body_25/pose_iter_584000.caffemodel rather than a caffemodel called pose_iter_XXXXXX.caffemodel (just rename your best model with that name, regardless of the number of iterations). No special flags are required to run it.

BODY_25B Model - Option 1 (Maximum Accuracy, Less Speed)

Thanks to Single-Network Whole-Body Pose Estimation, new models for BODY_25 were obtained with higher accuracy. We called it BODY_25B given that we add some extra PAF channels and body keypoints. It adds the MPII head and neck keypoints, and it removes the artifially created neck and middle hip of BODY_25 (that neck was simply the middle point of the shoulders and the middle hip the middle point of the hips).

To use the BODY_25B model trained for the Single-Network Whole-Body Pose Estimation paper (i.e., higher accuracy but slower speed than the default BODY_25 model):

  1. Download the Caffe model from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/1_25BSuperModel11FullVGG/body_25b/pose_iter_XXXXXX.caffemodel into experimental_models/1_25BSuperModel11FullVGG/body_25b/ as pose_iter_XXXXXX.caffemodel.
  2. Copy the body_25b folder into the OpenPose models folder, such as you end up with the previous 2 files in the following paths:
    1. openpose/models/pose/body_25b/pose_deploy.prototxt
    2. openpose/models/pose/body_25b/pose_iter_XXXXXX.caffemodel. Note: This model corresponds to pose_iter_464000.caffemodel (i.e., trained for 464k iterations), and it was renamed simply to follow the OpenPose format.
  3. Run the official OpenPose with your usual flags, while adding the --model_pose BODY_25B --net_resolution -1x480.
  4. Alternatively to 3, for multi-scale (even higher accuracy), then use --model_pose BODY_25B --net_resolution 1712x960 --scale_number 4 --scale_gap 0.25.

In addition, we also include all the files generated for and during training in experimental_models/1_25BSuperModel11FullVGG/training_results/. If you want to read the training_log.txt file, you can download it from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/1_25BSuperModel11FullVGG/body_25b/training_log.txt.

BODY_25B Model - Option 2 (Recommended)

We also include a second version of BODY_25B that is as fast as the default BODY_25, but that it highly reduces the number of false positives, increasing its overall accuracy.

To use this variation of the BODY_25B model trained for the Single-Network Whole-Body Pose Estimation paper:

  1. Download the Caffe model from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/1_25BBkg/body_25b/pose_iter_XXXXXX.caffemodel into experimental_models/1_25BBkg/body_25b/ as pose_iter_XXXXXX.caffemodel.
  2. Copy the body_25b folder into the OpenPose models folder, such as you end up with the previous 2 files in the following paths:
    1. openpose/models/pose/body_25b/pose_deploy.prototxt
    2. openpose/models/pose/body_25b/pose_iter_XXXXXX.caffemodel. Note: This model corresponds to pose_iter_464000.caffemodel (i.e., trained for 464k iterations), and it was renamed simply to follow the OpenPose format.
  3. Run the official OpenPose with your usual flags, while adding the --model_pose BODY_25B. Note that this model has been trained for the original and default --net_resolution -1x368, so this flag is not required.
  4. Alternatively to 3, for multi-scale (even higher accuracy), then use --model_pose BODY_25B --net_resolution 1312x736 --scale_number 4 --scale_gap 0.25.

In addition, we also include all the files generated for and during training in experimental_models/1_25BBkg/training_results/. If you want to read the training_log.txt file, you can download it from posefs1.perception.cs.cmu.edu/OpenPose/models/pose/1_25BBkg/body_25b/training_log.txt.