Voxelpose-with-visualization

About

We found 2 methods to visualize voxelpose-pytorch.

1. mmpose
2. Voxelpose itself

mmpose

1. requirements
   You may follow this installation guide
   
   
2. datasets
   Using mmpose to inference and visualize voxelpose only supports on CMU-Panoptic dataset yet. Prepare dataset by using panoptic-toolbox. We tested "160906_pizza1" and "170221_haggling_m3" and found these works well.
   
   
3. Run demo
   Please follow this docs to run demo. Add your datasets' path to run this demo.
   
   
4. Concat images and make video again
   We provide tools to concat image files(2d and 3d) and make those image files into demo video. Please refer to Scripts/concatimg.py, Scripts/mkvid.py. Add your own path to test this.
   
   
   

VoxelPose

1. requirements
   You may follow this installation guide
   For me torch version 1.7.1+cu110 worked well.
   
   
2. datasets
   Prepare the datasets by following this.
   We have trained on Campus dataset (epoch until 15 approximately), yet still show decent visualization.
   Please train more if wanted. Commands are same
   We have used the unfinished pretrained Shelf model. We have trained with three datasets('160224_haggling1','160226_haggling1','160422_ultimatum1') and validated with ''160422_haggling1''
   If you want to change the dataset for training, validation check ${POSE_ROOT}/lib/dataset/panoptic.py.
   You can get the pretrained models here. And place it in ${POSE_ROOT}
   
   
3. Run demo
   Download mkvid.py and visualize.py file from this repository then place the two files at ${POSE_ROOT}/Scripts
   We have made demo video on validation set if you want to visualize on different dataset, you can either simply modify the VAL_LIST given in ${POSE_ROOT}/lib/dataset/panoptic.py or modify the cofig file.
   After running one of the commands, demo_image file will be created forexample : ${POSE_ROOT}/output/panoptic/multi_person_posenet_50/prn64_cpn80x80x20_960x512_cam5/demo_image

python test/visualize.py -cfg configs/panoptic/resnet50/prn64_cpn80x80x20_960x512_cam5.yaml

python test/visualize.py --cfg configs/shelf/prn64_cpn80x80x20.yaml

python test/visualize.py --cfg configs/campus/prn64_cpn80x80x20.yaml

4. Concat images and make video again
   We provide tools to concat image files(located in demo_image) and make those image files into demo video. Please refer ${POSE_ROOT}/Scripts/mkvid.py. Add your own path to test this.

python Scripts/mkvid.py 

5. Precautions
   We have modified the ${POSE_ROOT}/lib/core/function.py issued in Shelf model.
   
   
   

results

mmpose

This is a demo video of "160905_pizza1" we have made. (only first 15 seconds uploaded)

VoxelPose

There are few videos we have made you can check it out

campus on shelf

Video	Dataset	Model
	Shelf	Campus_cam3
	Campus	Campus
	Shelf	Campus_cam5
	Shelf	Shelf
	Campus	Shelf_cam3
	Campus	Shelf_cam5
	Panoptic	Panoptic_cam5