Getting Started with DensePose

Inference with Pre-trained Models

Pick a model and its config file from Model Zoo(IUV), Model Zoo(CSE), for example densepose_rcnn_R_50_FPN_s1x.yaml
Run the Apply Net tool to visualize the results or save the to disk. For example, to use contour visualization for DensePose, one can run:

python apply_net.py show configs/densepose_rcnn_R_50_FPN_s1x.yaml densepose_rcnn_R_50_FPN_s1x.pkl image.jpg dp_contour,bbox --output image_densepose_contour.png

Please see Apply Net for more details on the tool.

Training

First, prepare the dataset into the following structure under the directory you'll run training scripts:

datasets/coco/
  annotations/
    densepose_{train,minival,valminusminival}2014.json
    densepose_minival2014_100.json   (optional, for testing only)
  {train,val}2014/
    # image files that are mentioned in the corresponding json

To train a model one can use the train_net.py script. This script was used to train all DensePose models in Model Zoo(IUV), Model Zoo(CSE). For example, to launch end-to-end DensePose-RCNN training with ResNet-50 FPN backbone on 8 GPUs following the s1x schedule, one can run

python train_net.py --config-file configs/densepose_rcnn_R_50_FPN_s1x.yaml --num-gpus 8

The configs are made for 8-GPU training. To train on 1 GPU, one can apply the linear learning rate scaling rule:

python train_net.py --config-file configs/densepose_rcnn_R_50_FPN_s1x.yaml \
    SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.0025

Evaluation

Model testing can be done in the same way as training, except for an additional flag --eval-only and model location specification through MODEL.WEIGHTS model.pth in the command line

python train_net.py --config-file configs/densepose_rcnn_R_50_FPN_s1x.yaml \
    --eval-only MODEL.WEIGHTS model.pth

Tools

We provide tools which allow one to:

easily view DensePose annotated data in a dataset;
perform DensePose inference on a set of images;
visualize DensePose model results;

query_db is a tool to print or visualize DensePose data in a dataset. Please refer to Query DB for more details on this tool

apply_net is a tool to print or visualize DensePose results. Please refer to Apply Net for more details on this tool

Installation as a package

DensePose can also be installed as a Python package for integration with other software.

The following dependencies are needed:

Python >= 3.7
PyTorch >= 1.7 (to match detectron2 requirements)
torchvision version compatible with your version of PyTorch

DensePose can then be installed from this repository with:

pip install git+https://github.com/facebookresearch/detectron2@main#subdirectory=projects/DensePose

After installation, the package will be importable as densepose.