LARD - YOLO V8

This repository contains material to train the YOLO v8 neural network architectures from Ultralytics on the LARD dataset, for detection, segmentation and pose estimation tasks.

Preview

Pose estimation

Palerme

Paphos

Detection

Palerme

Paphos

Segmentation

Palerme

Paphos

Setup

Example of installation:

conda create -p ./.conda python=3.10
pip install -r requirements.txt

NOTE: If necessary, you can override environment variables located in .env file by creating a .env.local one. It will automatically be loaded if existing.

Weights

All the neural networds are available under the nn directory tree, with ONNX exports and training associated files.

Metrics

Pose estimation

task	mult-adds (GFLops)	weights	mAP50 BBOX	mAP50:95 BBOX	mAP50 POSE	mAP50:95 POSE
pose	42.66	pretrained	0.99	0.9	0.98	0.95
pose	42.66	scratch	0.98	0.85	0.97	0.91

Detection

task	mult-adds (GFLops)	weights	mAP50 BBOX	mAP50:95 BBOX
detect	41.39	pretrained	0.99	0.91
detect	41.39	scratch	0.99	0.87

Segmentation

task	mult-adds (GFLops)	weights	mAP50 BBOX	mAP50:95 BBOX	mAP50 MASK	mAP50:95 MASK
segment	61.31	pretrained	0.98	0.87	0.97	0.76
segment	61.31	scratch	0.99	0.87	0.97	0.73