cruxes 0.1.18

A set of computer vision tools for analyzing your climbing videos.

Climbing Analysis Toolbox

A set of computer vision tools for processing and analyzing your climbing videos. In my spare time, I also write about topics relevant to bouldering and computer vision here.

Getting Started

# Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate

# Install or upgrade the published PyPI package
python -m pip install --upgrade pip
python -m pip install --upgrade cruxes

# Confirm the CLI is available
cruxes --help

The published package name is cruxes, and it installs the cruxes CLI.

PyPI: https://pypi.org/project/cruxes/

Catalogue

For each section, there will be detailed example code for both CLI usage and in-code usage.

1. Warping Video for Scene Matching Details

# Example usage:
cruxes warp \
--ref_img "examples/videos/warp-dynamic-ref.jpg" \
--src_video_path "examples/videos/warp-dynamic-input.mp4"
# [--type ...]

2. Drawing Trajectories for Body Movements [1] Details

cruxes body-trajectory \
--video_path "examples/videos/body-trajectory-input.mp4" \
--show_trajectory
# [other options]

3. Compare Body Trajectories across Different Climbing Footages Details

More to Come

• 3D Pose Extraction and Displaying
• Drawing Trajectories for Body Movements across Multiple Footages
• Heatmap for Limb Movement [1]
• Climbing Hold Auto-segmentation
• Gaussian-splatting 3D Reconstructing a Climb

---

1️⃣ Warping Video for Scene Matching

Sometimes, to analyze our sequences for a climb, we typically have multiple sessions. During those sessions, we might have the camera placed at different locations, thus pointing from different angles towards the climb we are projecting. This tool helps you transform videos so that they match a reference image that corresponds to the whole picture of your climb. Reasons for doing this are:

1. It is better for using tools that involve 2D/3D pose estimation
2. It is easier to see how your body moves with respect to similar angles. Note that, right now, it is impossible to seamlessly match a video to the scene of a base image if their camera angles and positions differ by a large amount; some area might be off from base scene.

To warp a video to match a reference scene, we extract the features between two frames, and then a homography matrix is extracted for the image transformation. By default, we use a per-frame homography matrix, but that also means we have to compute $H$ for each frame of the input video if the input video is moving. If the camera of your input video is not moving, we can reduce the processing time by only comparing the first frame of the video and the base scene. This reduces the computation time for the matcher we are using, so only image transformation is involved for the entire warping process. We call the first scenario dynamic and the second scenario fixed, as you can set with the type option.

# CLI usage
# Warp a video with moving camera (per-frame homography matrix for the transformation)
cruxes warp \
--ref_img "examples/videos/warp-dynamic-ref.jpg" \
--src_video_path "examples/videos/warp-dynamic-input.mp4"
# by default the type of warping is `dynamic`: `--type dynamic`

# In-code usage
from cruxes import Cruxes
cruxes = Cruxes()
cruxes.warp_video(
    "warp-dynamic-ref.jpg", 
    "warp-dynamic-input.mp4",
    # Optional: Advanced blending modes
    blend_mode="edge_feather",  # Options: 'none', 'feathered', 'edge_feather', 'smart', 'multiband', 'poisson'
    feather_amount=10,  # Pixels to feather at boundary (default: 10)
)

🎬 Example Resulting Video

Your browser does not support the video tag.

# CLI usage
# Warp a video with fixed camera (first-frame homography matrix for the transformation)
cruxes warp \
--ref_img "examples/videos/warp-fixed-ref.jpg" \
--src_video_path "examples/videos/warp-fixed-input.mp4" \
--type "fixed"

# In-code usage
from cruxes import Cruxes
cruxes = Cruxes()
cruxes.warp_video(
    "warp-fixed-ref.jpg", 
    "warp-fixed-input.mp4", 
    warp_type="fixed"
)

🎬 Example Resulting Video

Your browser does not support the video tag.

If you can't see the example resulting video, go to the example/videos/ folder.

---

2️⃣ Drawing Trajectories for Body Movements

It is recommended to apply this script to a video with fixed camera position, i.e., camera is not being moved.

There is a couple of settings you can adjust inside the script for extract_pose_and_draw_trajectory():

Argument	Description
track_point	Points of interest on the estimated pose you want to track. A velocity vector arrow will be drawn to indicate how fast each point is moving with respect to its 3D position
trajectory_only	Render only the trajectory on a black background. This disables pose drawing and telemetry, forces trajectory drawing on, and prefers cached trajectory metadata if available
overlay_mask	Whether to overlay a half-transparent mask on top of the original video.
hide_original_video	Whether to use a black background instead of the original video (useful for creating clean trajectory visualizations)
draw_pose	Whether to draw pose skeleton or not
pose_color	Color for the pose skeleton in BGR format (default: white (255, 255, 255))
show_trajectory	Whether to draw the trajectories (default: True)
show_gauges	Whether to show a top-left telemetry panel with raw_v and vel_ratio for each tracked joint
trajectory_history_seconds	If set, only the last N seconds of each joint trajectory are shown; if omitted, the full path is shown
use_cached_landmarks	Whether to reuse a matching landmarks JSON cache instead of recomputing pose landmarks
export_landmarks	Whether to save the collected pose landmarks to JSON after detection
landmarks_json_path	Optional cache file path. Defaults to <video_stem>_landmarks.json next to the input video
use_cached_trajectory_metadata	Whether to reuse a matching trajectory metadata JSON file as the trajectory source. This does not force drawing on by itself; show_trajectory still controls rendering
export_metadata	Whether to export unified frontend-facing metadata JSON, including per-sample displacement and per-second velocity vectors, per-frame pose landmarks, and explicit skeleton connections when pose data is available
metadata_path	Optional output path for the metadata JSON. Defaults to <video_stem>_trajectory_metadata.json next to the input video
kalman_settings	Whether to apply Kalman filter to smooth out the trajectory (not the pose itself)
savgol_settings	Whether to apply Savitzky-Golay filter to smooth the pose skeleton: [use_savgol, window_length, polyorder]
trajectory_png_path	Optional output path for a .png export of the trajectory on a black background

For CLI usage, --show_trajectory is required in the normal overlay mode. If you use --trajectory_only, trajectory drawing is enabled automatically.

--savgol_settings is currently available in the Python API example below, not in the CLI.

Then, run the command as follows:

# CLI usage
cruxes body-trajectory \
--video_path "examples/videos/body-trajectory-input.mp4" \
--trajectory_only \
--overlay_mask \
--draw_pose \
--show_trajectory \
--show_gauges \
--trajectory_history_seconds 2 \
--use_cached_landmarks \
--use_cached_trajectory_metadata \
--export_landmarks \
--export_metadata \
--kalman_settings 1e0
# Additional options:
# --hide_original_video  # Use black background
# --metadata_path ./my_metadata.json
# In trajectory_only mode, pose drawing and telemetry are disabled automatically.

# In-code usage
from cruxes import Cruxes
cruxes = Cruxes()
cruxes.body_trajectory(
    "body-trajectory-input.mp4",
    track_point=[
        # Currently available points to track
        "hip_mid",
        "upper_body_center",
        "head",
        "left_hand",
        "right_hand",
        "left_foot",
        "right_foot",
    ],
    trajectory_only=False,  # Set True for black-background trajectory-only output
    overlay_mask=False,
    hide_original_video=False,
    draw_pose=True,
    pose_color=(255, 255, 255),  # White color in BGR
    show_gauges=True,  # Show top-left telemetry for each tracked joint
    show_trajectory=True,
    trajectory_history_seconds=2.0,  # Show only the last 2 seconds; omit for full history
    use_cached_landmarks=True,  # Reuse a matching landmarks cache if present
    use_cached_trajectory_metadata=True,  # Reuse trajectory metadata for trajectory rendering if present
    export_landmarks=True,  # Save computed landmarks for later experimentation
    export_metadata=True,  # Export unified frontend-facing metadata JSON
    kalman_settings=[  # Kalman filter settings: [use_kalman : bool, kalman_gain : float]
        True,  # Set this to false if you don't want to apply Kalman filter
        1e0,  # >=1e0 for higher noise, <=1e-1 for lower noise
    ],
    savgol_settings=[  # Savitzky-Golay filter: [use_savgol, window_length, polyorder]
        True,  # Set to True to smooth pose skeleton
        15,  # Window length (must be odd, typical: 5-15)
        4,  # Polynomial order (typical: 2-4, must be < window_length)
    ],
    trajectory_png_path=None,
)

The generated video will be saved in the same directory as your input video with a pose_trajectory_ prefix.

🎬 Example Resulting Video

Your browser does not support the video tag.

If you can't see the example resulting video, go to the example/videos/ folder.

---

3️⃣ Compare Body Trajectories across Different Climbing Footages

To be added.

---

To-do

• Add automated test cases
• Add specification to notice for adding new tool kits in the future
• Add a server backend to allow API request for specific functionality.
• Minimize pose estimation to unit functions and apply Kalman filter by default to smooth out the jiggling.
• Migrate to PyPI for easier installation and use.
• Add CLI options to run (cruxes instead of python ...)