JcProg/pedsense-yolo11m-pose-jaad-10e

PedSense -- YOLO11m-Pose Pedestrian Keypoint Model (JAAD, 10 epochs)

Overview

A pedestrian keypoint detection model fine-tuned on the JAAD (Joint Attention in Autonomous Driving) dataset using the PedSense-AI framework. Detects pedestrians and predicts 17 COCO body keypoints per person.

Keypoint labels were generated by running the pretrained yolo11n-pose model on JAAD dashcam frames (pseudo-labelling). The model was then fine-tuned from yolo11m-pose on those labels.

Model Details

Property	Value
Architecture	YOLO11m-Pose
Task	Keypoint detection (1 class: pedestrian, 17 COCO keypoints)
Dataset	JAAD
Epochs	10
Image Size	640x640
Batch Size	16
Train/Val Split	80/20 (video-level)
kpt_shape	[17, 3]

Performance (Best checkpoint -- epoch 9)

Box metrics

Metric	Value
Precision	0.876
Recall	0.945
mAP@50	0.873
mAP@50-95	0.962

Pose metrics

Metric	Value
Precision	0.840
Recall	0.899
mAP@50	0.821
mAP@50-95	0.900

COCO Keypoint Order (17 points)

Index	Keypoint
0	Nose
1	Left eye
2	Right eye
3	Left ear
4	Right ear
5	Left shoulder
6	Right shoulder
7	Left elbow
8	Right elbow
9	Left wrist
10	Right wrist
11	Left hip
12	Right hip
13	Left knee
14	Right knee
15	Left ankle
16	Right ankle

Usage

from ultralytics import YOLO

model = YOLO("weights/best.pt")
results = model("image.jpg")

# Access keypoints
for r in results:
    print(r.keypoints.xy)   # pixel coordinates (N, 17, 2)
    print(r.keypoints.conf) # confidence per keypoint (N, 17)

Training

Keypoint labels were generated with preprocess pose, then the model was trained using the PedSense-AI framework:

# Step 1: Extract frames
uv run pedsense preprocess frames

# Step 2: Generate pose labels (pseudo-labelling with yolo11n-pose)
uv run pedsense preprocess pose

# Step 3: Train yolo11m-pose on generated labels
uv run pedsense train -m yolo-pose -n my_pose_model -e 10 -b 16 --yolo-variant yolo11m-pose

License

AGPL-3.0

Related Models

• pedsense-yolo26m-detector-jaad-20e -- 1-class pedestrian detector (20 epochs)
• pedsense-yolo26m-detector-jaad-aug-10e -- augmented detector (10 epochs)
• pedsense-yolo26-jaad-10e -- 2-class crossing intent classifier