Run script for "Inference in the wild" of https://github.com/facebookresearch/VideoPose3D - Make sure requirements are met, described in comment

run_videopose3d__in_the_wild.sh

#!/bin/sh

# The input videos to detect - if directory apply script on each video of it  - use absolute path!
INPUT_PATH="/path/to/your/videos/"

# Path were detected annotations and rendered videos are stored  - use absolute path!
OUTPUT_DIR="$HOME/Development/VideoPose3D/output_directory"

# Custom name for the generated dataset - here: using name of directory $INPUT_PATH
#   Used to generate data/data_2d_custom_<DATASET_NAME>.npz
DATASET_NAME=$(basename "$INPUT_PATH")

# Step 1: setup
#########################################################
echo 'STEP 1: setup...'

# Download the pretrained model for generating 3D predictions.
if [ ! -f checkpoint/pretrained_h36m_detectron_coco.bin ]; then
  wget -O checkpoint/pretrained_h36m_detectron_coco.bin https://dl.fbaipublicfiles.com/video-pose-3d/pretrained_h36m_detectron_coco.bin
fi
echo ' > STEP 1: ✅'


# Step 2 (optional): video preprocessing
#########################################################
echo 'STEP 2: Skipping optional video preprocessing...'

# Cut start and end
#ffmpeg -i input.mp4 -ss 1:00 -to 1:30 -c copy output.mp4

# Optionally adapt the frame rate of the video - our Human3.6M model was trained on 50-FPS videos
#ffmpeg -i input.mp4 -filter "minterpolate='fps=50'" -crf 0 output.mp4
echo ' > STEP 2: ✅'


# Step 3: inferring 2D keypoints with Detectron
# Using Detectron2 (new)
#########################################################
echo 'STEP 3: inferring 2D keypoints with Detectron...'

mkdir -p "$OUTPUT_DIR"  # Create output dir if not exists

cd inference
python infer_video_d2.py \
    --cfg COCO-Keypoints/keypoint_rcnn_R_101_FPN_3x.yaml \
    --output-dir "$OUTPUT_DIR" \
    --image-ext mp4 \
    "$INPUT_PATH"
cd ..

echo ' > STEP 3: ✅'


# Step 4: creating a custom dataset
#########################################################
echo 'STEP 4: creating a custom dataset...'

cd data
python prepare_data_2d_custom.py -i "$OUTPUT_DIR" -o "$DATASET_NAME"
cd ..

echo ' > STEP 4: ✅'


# Step 5: rendering a custom video and exporting coordinates
#########################################################
echo 'STEP 5: rendering a custom video and exporting coordinates...'

# Render ALL videos with pose information baked into
for video in "${INPUT_PATH}"*.mp4
do
  name=$(basename "$video")
  echo " Rendering $name"
  python run.py \
      -d custom \
      -k "$DATASET_NAME" \
      -arc 3,3,3,3,3 \
      -c checkpoint \
      --evaluate pretrained_h36m_detectron_coco.bin \
      --render \
      --viz-subject "$name" \
      --viz-action custom \
      --viz-camera 0 \
      --viz-video "$video" \
      --viz-output "${OUTPUT_DIR}/${name}__out.mp4" \
      --viz-size 6
done

echo ' > STEP 5: ✅'

Run script for facebookresearch / VideoPose3D
❗Make sure you installed the dependencies: ❗

• facebookresearch / VideoPose3D
  • Python 3+ distribution
  • PyTorch >= 0.4.0
• facebookresearch / Detectron2 (required by VideoPose3D for "Inference in the wild"): Detectron2 install instructions