Home Models Journal About GitHub
DOMAIN

Autonomous Perception

Real-time visual understanding for machines that navigate the world. Object detection, drivable area segmentation, and monocular depth estimation trained on BDD100K driving data.

THE PIPELINE
๐Ÿš—
Object Detection

Real-time detection of vehicles, pedestrians, and traffic signs using transformer-based architecture.

RT-DETR BDD100K Fine-tuned
๐Ÿ›ฃ๏ธ
Drivable Segmentation

Pixel-level drivable area classification for safe navigation using encoder-decoder architecture.

U-Net BDD100K Custom
๐Ÿ“
Depth Estimation

Monocular depth inference from a single camera frame for distance estimation.

DPT Transformer Real-time
MODEL PERFORMANCE
~0.45
Detection mAP@50
RT-DETR on BDD100K
~0.75
Segmentation IoU
U-Net Drivable Area
Real-time
Inference Speed
All models combined
LIVE DEMO
Autonomous Perception Ensemble
Open in new tab โ†’
RT-DETR U-Net DPT BDD100K PyTorch Gradio Hugging Face
DATASET
BDD100K

Berkeley Deep Drive dataset with 100K driving videos and rich annotations for detection, segmentation, and drivable area classification across diverse weather and lighting conditions.

100K
Videos
10
Tasks
8
Categories