Skip to content

What is SLEAP?

SLEAP (Social LEAP Estimates Animal Poses) is a deep learning framework for multi-animal pose estimation. While designed for tracking animal body parts, SLEAP works exceptionally well for plant root phenotyping.

Why SLEAP for Roots?

SLEAP excels at identifying keypoints in images and learning their structural connections. For plant roots, this means the neural network directly outputs a graph representation – nodes (landmarks) and edges (skeleton connections) – for each root, with multiple roots grouped per plant.

Key Advantages for Root Phenotyping

What makes SLEAP powerful for roots is the direct extraction of structured graphs from images:

  • No post-processing – Get root skeletons directly from the neural network output
  • Efficient trait computation – Any geometric trait can be constructed from the extracted nodes and edges
  • Multi-instance detection – Detect multiple roots simultaneously in each frame
  • Temporal tracking – Associate roots across frames (works best with minimal occlusions; may require proofreading)
  • Flexible architectures – Define custom skeletons for different root types (primary, lateral, crown)
  • Fast inference – GPU-accelerated prediction scales to large datasets

SLEAP Workflow

The typical SLEAP workflow for root phenotyping:

graph LR
    A[Root Images] --> B[Label Landmarks]
    B --> C[Train Model]
    C --> D[Run Inference]
    D --> E[SLEAP Predictions .slp]
    E --> F[sleap-roots Analysis]
    F --> G[Trait CSV]
  1. Collect images – Time-lapse or single time-point root images
  2. Label landmarks – Annotate key points (tips, bases, nodes) in SLEAP GUI
  3. Train model – Use SLEAP to train a pose estimation model
  4. Run inference – Generate predictions (.slp files) for all images
  5. Analyze with sleap-roots – Extract morphological traits from predictions

Root Skeleton Design

A SLEAP skeleton defines the structure of landmarks to track. For roots, we typically use:

Dicot Skeleton (Primary + Lateral Roots)

Primary Root:
  base → node1 → node2 → ... → tip

Lateral Roots (multiple instances):
  base → node1 → node2 → ... → tip

Monocot Skeleton (Crown Roots)

Crown Roots (multiple instances):
  base → node1 → node2 → ... → tip

Each root is a connected chain of landmarks from base to tip.

Prediction Files (.slp)

SLEAP generates .slp files containing:

  • Landmark coordinates (x, y) for each point
  • Instance tracks – Which points belong to which root
  • Confidence scores – Model certainty for each prediction
  • Frame associations – Temporal tracking across frames

sleap-roots uses the sleap-io library to load and parse these files.

Example: Loading SLEAP Predictions

import sleap_io as sio

# Load predictions
labels = sio.load_slp("primary_roots.slp")

# Access first frame
frame = labels.labeled_frames[0]

# Get first instance (root)
instance = frame.instances[0]

# Get landmark points
for i, node in enumerate(instance.skeleton.nodes):
    point = instance.points[i]  # point is a structured array with coordinates, score, visibility
    x, y = point[0][0], point[0][1]  # point[0] is the [x, y] coordinate array
    print(f"{node.name}: ({x:.1f}, {y:.1f})")

Output: 

r1: (1016.8, 144.4)
r2: (1208.0, 304.1)
r3: (1208.9, 472.4)
...
r6: (1136.1, 1021.0)

From SLEAP to Traits

sleap-roots converts SLEAP predictions into trait measurements:

import sleap_roots as sr

# Load SLEAP predictions
series = sr.Series.load(
    series_name="my_plant",
    primary_path="primary_roots.slp",
    lateral_path="lateral_roots.slp"
)

# Extract points for a frame
pts = series.get_primary_points(frame_idx=0)
# pts is a (n_instances, n_nodes, 2) numpy array of (x, y) coordinates

# Compute length from points
from sleap_roots.lengths import get_root_lengths
length = get_root_lengths(pts)
print(f"Primary root length: {length:.2f} pixels")

The Series class handles loading, organizing, and extracting point data from SLEAP files.

Data Organization

Single Root Type

For experiments with only one root type (e.g., primary only):

my_plant.slp  # SLEAP predictions
my_plant.h5   # Optional: original video

Load with: 

series = sr.Series.load(
    series_name="my_plant",
    primary_path="my_plant.slp",
    h5_path="my_plant.h5"  # Optional
)

Multiple Root Types

For dicots (primary + lateral) or monocots (primary + crown):

my_plant.primary.slp  # Primary root predictions
my_plant.lateral.slp  # Lateral root predictions
my_plant.h5           # Optional: video

Load with: 

series = sr.Series.load(
    series_name="my_plant",
    primary_path="my_plant.primary.slp",
    lateral_path="my_plant.lateral.slp",
    h5_path="my_plant.h5"
)

Batch Experiments

For high-throughput experiments:

experiment/
  plant1.primary.slp
  plant1.lateral.slp
  plant1.h5
  plant2.primary.slp
  plant2.lateral.slp
  plant2.h5
  ...

Use auto-discovery: 

paths = sr.find_all_slp_paths("experiment/")
plants = sr.load_series_from_slps(paths, h5s=True)

SLEAP vs sleap-roots

Tool Purpose Output
SLEAP Pose estimation Landmark coordinates (.slp)
sleap-roots Trait extraction Morphological measurements (CSV)

SLEAP is upstream – it detects and tracks root landmarks. sleap-roots is downstream – it computes biological traits from those landmarks.

Learning SLEAP

To get started with SLEAP for root phenotyping:

  1. Official SLEAP Docs – https://sleap.ai/tutorials/
  2. Video Tutorial – SLEAP GUI walkthrough on YouTube
  3. Example Projects – Use sleap-roots test data as reference
  4. Community – SLEAP Discussions

Best Practices for Root Labeling

When creating SLEAP training data:

  • Label consistently – Always start at base, end at tip
  • Label enough frames – 50-100 frames is usually sufficient
  • Use suggestion mode – Let SLEAP's active learning guide labeling
  • Label all instances – Include all roots in each frame
  • Verify predictions – Check inference quality before large-scale runs

Coordinate Systems

SLEAP predictions use image coordinates:

  • Origin (0, 0) is top-left corner
  • X-axis increases rightward
  • Y-axis increases downward

sleap-roots respects this coordinate system. For angle calculations, we account for the inverted Y-axis.

Next Steps