Skip to content

Latest commit

 

History

History
162 lines (122 loc) · 4.44 KB

README.md

File metadata and controls

162 lines (122 loc) · 4.44 KB
Raw

Before training the models you need to download the ImageNet pre-trained ResNet weights

$ cd models/pretrained
$ ./download.sh

Training parameters are specified in the pose_cfg.yaml file.

Here are the dataset specific instructions.

Training a model with MPII Pose Dataset (Single Person)

  1. Download the dataset from this page, both images and annotations. Unpack it to the path <path_to_dataset> to have the following directory structure:
<path_to_dataset>/images/*.jpg
<path_to_dataset>/mpii_human_pose_v1_u12_1.mat
  1. Preprocess dataset (crop and rescale)
$ cd matlab/mpii
$ matlab -nodisplay -nosplash

# in matlab execute the following function, be sure to specify the *absolute* path
preprocess_single('<path_to_dataset>')
  1. Edit the training definition file models/mpii/train/pose_cfg.yaml such that:
dataset: `<path_to_dataset>/cropped/dataset.mat`
  1. Train the model
$ cd models/mpii/train/
$ TF_CUDNN_USE_AUTOTUNE=0 CUDA_VISIBLE_DEVICES=0 python3 ../../../train.py

Training a model with MS COCO dataset (Multi-Person)

  1. Download MS COCO train2014 set with keypoint and object instances annotations.

  2. Download pairwise statistics:

$ cd models/coco
$ ./download_models.sh
  1. Edit the training definition file models/coco/train/pose_cfg.yaml such that:
dataset: `<path_to_mscoco>`
  1. Train the model:
$ cd models/coco/train/
$ TF_CUDNN_USE_AUTOTUNE=0 CUDA_VISIBLE_DEVICES=0 python3 ../../../train.py

Training on your own dataset

If you wish to train keypoint detectors on your own data, first of all you have to prepare dataset definition file dataset.mat (Matlab data format). We included an example of such file here. It must contain a variable dataset that is a struct array, with each entry corresponding to an image and with the following fields:

  1. image - path to the image
  2. size - 1x3 array containing [num_channels, image_height, image_width]. Set num_channels=3 for an RGB image.
  3. joints - a cell array of nx3 joint annotations, for example:
joints = {[ ...
  0,  175,  261; ... % 0-indexed joint ID, X coordinate, Y coordinate
  1,  173,  178; ...
  2,  144,  122; ...
  3,  193,  124; ...
]};

You will potentially need to adjust the training definition file pose_cfg.yaml. Here we will describe some of its options:

# path to the dataset description file
dataset: /path/to/dataset.mat

# all locations within this distance threshold are considered
# positive training samples for detector
pos_dist_thresh: 17 

# all images in the dataset will be rescaled by the following
# scaling factor to be processed by the CNN. You can select the
# optimal scale by cross-validation
global_scale: 0.80

# During training an image will be randomly scaled within the
# range [scale_jitter_lo; scale_jitter_up] to augment training data,
# We found +/- 15% scale jitter works quite well.
scale_jitter_lo: 0.85
scale_jitter_up: 1.15

# Randomly flips an image horizontally to augment training data
mirror: true

# list of pairs of symmetric joint IDs, for example in this case
# 0 and 5 are IDs for the symmetric parts, and 12 or 13 do not have
# symmetric parts. This is used to do flip training data correctly. 
all_joints = [[0, 5], [1, 4], [2, 3], [6, 11], [7, 10], [8, 9], [12], [13]]

# Type of the CNN to use, currently resnet_101 and resnet_50
# are supported
net_type: resnet_101
init_weights: ../../pretrained/resnet_v1_101.ckpt

# Location refinement parameters (check https://arxiv.org/abs/1511.06645)
location_refinement: true
locref_huber_loss: true
locref_loss_weight: 0.05
locref_stdev: 7.2801

# Enabling this adds additional loss layer in the middle of the ConvNet,
# which helps accuracy.
intermediate_supervision: true
intermediate_supervision_layer: 12

# all images larger with size
# width * height > max_input_size*max_input_size are not used in training.
# Prevents training from crashing with out of memory exception for very
# large images.
max_input_size: 850

# Learning rate schedule for the SGD optimiser. 
multi_step:
- [0.005, 10000]
- [0.02, 430000]
- [0.002, 730000]
- [0.001, 1030000]

# How often display loss
display_iters: 20

# How often to save training snapshot
save_iters: 60000

You don't have to crop images such that they all have the same size, as training is done with batch_size=1 (batch size larger than 1 is currently not supported anyway).